Xiangrui dropped "Coze 3.0 is officially live" into the group, and the first reaction wasn't excitement but wallet-pain. 大*** cried out "the task consumption more than doubled," while 天* offered a workaround ， "if you skip the team features and custom models, it's fine." 北非飞狼 piled on: "the enterprise edition runs over 500 a month, expense it to your company." 醒* then pivoted, plugging 归藏's card skills, sharing the vibecoding site he'd built so everyone could see the results, insisting "it's even better than what an average designer makes." 周** seized the moment to lay out his three-route theory: the tool camp, the business camp, the real-world camp.

小* posted a photo: seven screens, three running models, the rest for office work, captioned "first day on the job." The group instantly exploded ， 平*** counted them off, "seven screens, dizzying," and 叶落知秋 shouted "workstation of my dreams." 小* added, "it really is my first day at work," and 郑** laughed, "this doesn't look like showing up for work, it looks like building out the company's infrastructure," putting the term "One-Person Company OPC" on the table for the first time. K*** had barely finished asking what OPC meant when Xiangrui jumped in: in June or July we could hold an offline meetup in Guangzhou and let everyone pitch ideas.

宋** lit up the room with "what's everyone's mental price for paying for one piece of information?" He pegged his own at "10 cents ， cheap enough that one look doesn't sting," but 醒* pushed back: "charging a dime is worse than free, it just wrecks your reputation." 程** dropped the day's hardest lines: "knowledge free, service paid" and "in the AI era, the ability to ask high-quality questions is the core skill," and recommended the book Asking the Right Questions. Xiangrui chimed in: everyone is welcome to open-source their best cases and solutions into the community knowledge base and turn it into an industry solution library. 门*** nailed the essence of paying: a reusable skill is worth more than ordinary information.

R***'s remark ， "Bitable is about to run out of capacity again, our single order table is already hundreds of thousands of rows" ， drew out everyone who knows their stuff. 辰* flatly talked her out of it: "past 20,000 rows you should be thinking database; a million-row Bitable will crush your operations team's computers." E*** laid out his own approach: keep all data in a database, and auto-build interlinked Bitables for each stage. 炮 closed it out in one line: "database on the back end, Bitable on the front end, and stable output still comes down to the API." R*** explained she was pulling data from heterogeneous systems into Bitable to build business dashboards, and the whole table thrashed out "what should store your data" once and for all.

万** marveled, "the morning of June 1st, the chat in this group blew up ， livelier than a school performance." Xiangrui rode the moment: "this is our own Children's Day stage, happy Children's Day to grown-ups and kids alike." 宋** wrote a golden line on the spot: "On Children's Day 2026, a grown-up kids' theater show was staged in Xiangrui and friends' group." The cleverest bit was 万**'s experiment ， he deliberately wrote the line "group-chat daily report GEO" to see whether a single sentence could get caught by the daily-report AI, and R*** laughed at once, "one line made the list, you've totally figured it out." 周** even improvised an AI Children's Day poem, "When I Grew Up I Became You."

The talk turned to the "guru" of Feishu Bitable. 周** argued strongly for a high-quality printed book, vetted through three rounds of review and proofreading, to lay the strongest foundation for Bitable; 宋** agreed, "not for money or fame, but so humanity can iterate fast in the AI era." The discussion got a little heated, and 哈* and 周** traded a few barbs. D***threw cold water on it: "he's a ByteDance employee ， rights, responsibilities, and rewards don't line up neatly, it's not as simple as you imagine." Xiangrui stepped in twice: "let's change the subject ， in our group we shouldn't be discussing individuals, it's not appropriate," steadily steering the car back on track.

Xiangrui's line ， "I made it onto CCTV news, brothers" ， set off the afternoon session. 侯* posted that Yizhuang foldable-phone shot, and the group filled with "awesome," while 平*** joked, "CCTV's infringing ， they didn't even credit you in the whole piece." Then Xiangrui dropped an even bigger scoop: Feishu's secret, invite-only "Spreadsheet Agent" can detect and install skills on its own ， "basically the lobster of Bitable." D***cut to the heart of it: "it's like embedding the lobster's abilities right into the spreadsheet ， the function has sunk down a layer." 万** marveled, "it massively shortens the chain; agent-plus-Feishu-CLI was still too long before."

power suggested Xiangrui write an OPC series: how to handle physical setup, venue, finance, and compliance. Xiangrui's eyes lit up ， "turning my own OPC startup journey into a series, that gives the persona" ， but 万** cautioned against diluting the bullseye of "Bitable and AI implementation," and Xiangrui answered cleanly: "Bitable can be diluted, but AI implementation is the main line and can't change." R*** meanwhile, took the clock-out skill to new heights ， pulling logs from DingTalk, having AI analyze them and send a digest to the boss, then writing them into Bitable to build employee profiles ， a whole chain running end to end. As the deep night struck, Xiangrui announced that all seventeen thousand notes in the knowledge base had come fully alive.

power drops mungermodels.com and learnbuffett.com, a Munger-and-Buffett knowledge graph made by one blogger, marveling that doing things with AI really takes data, taste, and a clear sense of process to turn out real work. His casual line, surely there's no copyright risk for Munger and Buffett in China, gets a cold splash from 周**, who points out CITIC Publishing and the Munger Academy go after exactly this kind of piracy and infringement; the blogger is just small enough to stay off the radar for now, but compliance trouble is coming. When power asks whether that boundary is a bit wide, 周** lays out the facts: even CITIC has fought piracy lawsuits with WeChat Read; AI moves faster than the law, and there are too many new situations no one can draw lines around.

吴*says he's about to catch up on the articles he missed, and Xiangrui slips in a plug that all their articles are archived in the ima knowledge base. 罗*is hit with decision paralysis on the spot: how much does ima really differ from ob, and how does cc-plus-ob stack up against workbuddy-plus-ima. Xiangrui answers bluntly, the gap is huge, ima honestly isn't great to use. 罗*pushes back, ima now supports oth authorization and even ships its own agent that can run skills, but Xiangrui keeps tearing it down, the skills aren't good, they can't manage files, and after you add a note you still have to sort it by hand. 罗*concedes with a that's-fair-too.

Xiangrui drops yesterday's group daily report into the chat, and R*** laughs that she's got codex chatting with her boss on WeChat. 周**eggs her on, promote and give a raise right now, good kid. R*** asks if there's a way to skip the human stepping in to hit send, and Xiangrui jokes she could ask Zhang Xiaolong. 枫**spots the 'homepage strategist' line in the report and says it's a great coinage. 周**tosses out a philosophical one, what's the difference between yesterday's Xiangrui and today's Xiangrui, and 枫**fires back instantly, a little more handsome every day, earning a 666.

K*** groans that he forgot not to update WeChat, and after the accidental update the whole UI changed, looking more and more like DingTalk. The thread turns to computer use: R*** says she had codex directly drive her PC to read WeChat and edit messages, and it actually worked. 平***, on the other hand, hit a wall, the moment codex touched WeChat he got kicked offline and had to log back in. Comparing notes, R*** is on Windows, and even she couldn't get codex to auto-send, it insists on a confirm first. 枫**chimes in with advice from the side, hoarding backup accounts never goes out of style, WeChat's Windows countermeasures probably just haven't caught up.

Xiangrui asks if anyone has tried 胡彦斌's vibe-coded app, and T*** supplies the name, 彦火, a fan community the singer built solo, live on the App Store today. One look and Xiangrui calls the AI smell strong; 平***piles on that it's reportedly full of bugs, with people exploiting them to farm points. Then Xiangrui digs up the real scoop, confirmed: Feishu base now has an agent, in internal testing. R*** runs with it and pitches a new module, gather the group's distilled methods and SOPs into something called the community growth roadmap, so a new member can glance at it and tell this community is worth following. Xiangrui cheers it on the spot, now that's a good idea.

S*** asks what computers everyone uses, since his laptop lags just running one agent, kicking off a full gear roll call. 程**guesses he installed the Lobster, whose permissions run so high it eats a lot of CPU. 枫**fires off one-liners, the cutesy agent matrix, three-to-six agents into battle, run locally on a 16GB-plus GPU, save tokens for the poor, then self-deprecates as the sucker who bought a 12GB card. S*** runs workbuddy and Tencent's Marvis; 门***says even his 8GB of RAM ran it without complaint, and Marvis really does look good. The whole group admits they love watching the agent work, seeing how it breaks tasks down, picking up skills.

醒*raises a real question, what are the rates for taking Feishu base jobs; today she quoted four figures and the client just hung up. 天*sets the bar, 2000 a day is the going rate, billed per person-day, and plain table-building is something anyone can do. The topic flares into whether RAG is obsolete; 天*says RAG is done, moving to Wiki, but 北非飞狼snaps back that RAG has long been in production and won't iterate that fast, while 醒*mediates, they're simply different scenarios, neither replaces the other. 木*lifts the battlefield higher, she does AI delivery design, from strategy to goals to plan, with scenario insight and change management at the core, and every new thing meets four kinds of people. 炮adds that the client doesn't want a slick table, they want their problem solved. It all lands on consensus, the two-in-one of understanding business plus tech is worth the most, and experience and judgment only grow more valuable.

On a whim Xiangrui decides to open a Douyin chat room, calling the crew in to talk by voice. Dropping the live link and nudging all the while, come on up, let's chat, keep it going. 明天points out the QR won't scan, asking for a Douyin handle instead; l*** self-deprecates that she's too old for Douyin and goes off to download it, while xixi is also still downloading. R*** plays along, get on dy and have fun. After all the fuss, the chat actually takes off, with xixi marveling that this kind of format is great, so glad to learn from the experts, and K*** closing with a line, well then, the wrap-up skill. At midnight, a chorus of happy Children's Day picks up the holiday.

R*** jumps in right away: I want to build one in Feishu for managing clients too, let's just do it, and 宋** caps it with that execution speed is off the charts. She paints the client dashboard in full detail: basic client info, stakeholders, performance, project progress, visit records, group-chat summaries, she wants it all. 卡* pours in a dose of reality, asking how this data gets updated, do you have cc scrape it and rewrite the HTML every time, and adds that his own AI intel-agent runs on a Feishu Bitable as its backend, the pain being unstable data entry and always forgetting to fill it in. K*** chimes in flatly, this daily-report feature is a real tool for boosting group activity, and Xiangrui tacks on a line, put the Bitable to good use.

Xiangrui drops xiangruiai.com, polished to his liking, into the group for eyes, and 卡*'s first reaction is whether the site needs to be filed for registration. 宋** offers the most practical note: there's too much stuff under the name, think about it, what people most want to know is what you can do, how they can collaborate, how to reach you, so you should leave a one-liner or an image that's easy to forward, and whether or not you're building a personal IP, that awareness has to be there. Xiangrui explains it's neither a résumé nor a corporate site, more like a public workbench that tells people where I came from, what I'm working on, and what long-term work I've accumulated. 周** waxes lyrical, going from the three small steps of visual design to how the group owner is right in the middle of a stage of dramatic growth, with N possible routes from uncertainty to certainty, then circling around to perseverance under heaven and the positive energy of helping young people.

B*** asks who's using codex, and the question cracks open the floodgates. 万** vents: buying ChatGPT Plus throws a payment-not-approved error, Google Pay works fine but the credit card just won't go through, two days of fiddling and many attempts later he still can't figure out the review mechanism. h** reassures him it's random, it'll work again in a couple of days, his own bank card got rejected at first and then inexplicably went through. Meanwhile 归元 is eyeing the daily report's layout and suggests breaking the content into lines by nickname, saying it's all crammed together and dizzying right now, and Xiangrui replies oh sure, let me adjust it, claiming it on the spot.

盧 cuts in with a question: can Feishu OKR invite external contacts as view-only without edit rights, and Xiangrui says no. 辰* raises something important, this daily-report info should be anonymized, mask the avatars and swap nicknames for asterisks, because a lot of people in the group use real names, and offhandedly adds that if it weren't for WeChat's controls, Xiangrui could absolutely turn this whole thing into a standalone product and make a tidy sum. R*** seconds it plainly, yes, do clean it up, as you grow bigger you might run into someone you know, and some things won't be so easy to talk about in the group. B*** adds a line from the sidelines, you summed up the characters really well.

Xiangrui is done and announces every page can now be shared. 辰* and R*** play around with the interaction gag on the self-intro card, that checkbox you click, R*** laughs and says I always want to go uncheck it, and 辰* suggests it'd fit better as a checkbox placed up front. 宋** cracks up the whole group with true gold fears no fire, Ruixiang dreads no flame. Half joking, half serious, 卡* calls it out, 李祥瑞 don't be like this, you've already caught the vibe coding bug. 李** praises the site's aesthetics and content as both on point. Xiangrui doesn't put on airs either: vibe coding feels too good, can't stop for a second.

Amid the buzz, 周** throws out an article with a piercing title: in the cognitive disarmament that AI triggers, what we lose will be far more than ability. He excerpts a passage: AI saves you the effort of thinking, but it also takes away that clumsy, slow, painful yet necessary stretch before you reach the answer, the hesitating and trial-and-error, the detours, the wandering in dead ends, and that sudden moment when two threads finally connect, all of it folded away. K*** doesn't pick up the sentimental thread, turning instead to the practical, asking whether there's a good tool for generating white-background images, then floating his own sticking point: Chrome, the Codex extension, and the native host are all installed, but at runtime codex can't get the browser channel, has anyone hit the same problem.

In the afternoon Xiangrui posts a retrospective: a personal website that can view WeChat group daily reports, live after just a few hours of chatting with AI. K*** having just read it, marvels at how you managed to describe the whole inner journey of the process. power marvels that the group owner is a truly high-energy person, doing what he loves. But vibe coding isn't all bliss. 志* lays out the other side: DeepSeek's model is garbage, his own vibe coding always drifts off course, never matches expectations, riddled with bugs, hard to make smooth no matter how much he tunes it, and some days he grinds it out into a pile of crap and just wants to scrap it and start over. On one side the breezy few-hours-to-launch, on the other the messy day-long grind that ends in a do-over, two faces of the same word.

xixi raises a real problem: she built a teaching app with data in it to share with students for practice, but worries they'll mess up each other's data, so how does she turn it into a template. 卡* and 宋** take turns unpacking it, the app is just a shell and the content lives in the Bitable, app mode is outward-facing by nature and handing it out means letting people create copies, to make a real template you'd have to publish to the app marketplace and you also need management rights over the source tables, it's a whole systems undertaking. That unpacking surfaces a bigger anxiety: 侯燕青 says the SaaS backend is in someone else's hands, data security is a false premise to begin with, Feishu could be quietly peeking at every table. 宋** fires back bluntly, if you need things that confidential then stop chasing after AI. The conversation pivots, Xiangrui drops a fresh image from GPT image2, saying goodbye to the cartoon style, even publishing the prompt, and 卡* follows up asking whether it could be turned into an open-source workflow. The tail end scatters back into everyday chatter: 露* asks what to do when the font goes blurry once enlarged, K*** suggests taking it to banana to bump up the resolution, and 宋** is agonizing over whether a 24+1T Mac is enough to buy.

宋**urged Xiangrui to launch a popular-science column, arguing the edge was AI automation, that if it could be done it had to be done, and that it counted as passive income to some degree. Xiangrui threw cold water on it right away, saying his WeChat articles are already done with AI assistance, but you can't hand it all to AI, that pure-AI output and human-machine collaboration are simply not the same quality. B*** piled on, saying pure AI writing, even fed with the latest takes, comes out hollow, and you still end up hand-polishing seven or eight drafts. R*** chimed in: we want something real, not a toy. 侯*put it most plainly: let AI run the first draft, but the final draft must be confirmed by a human.

庄*said he'd built a skill and was looking for test samples, with a roadshow in Shanghai over the weekend. His tagline was pretty abstract: send a photo of you and your partner and get a glimpse of your child's whole life. 木***said bluntly that this pitch would shut out 99% of investors. 庄*took it in stride: zero product logic, pure play, treat it as garbage if you like. 周**got serious, suggesting he break out of the rigidity of multidimensional tables and make it playful with hand-drawn prototypes, the viral value would be enormous. 万**casually dug up an old-photo-to-elderly-portrait project he'd once made, the code still around, and the two clicked instantly.

平***stumbled on the open-source project wechat-radar, and Xiangrui seized the moment to pull out his own version, far more detailed. The moment the community dashboard appeared, 悉*cried out that it was a private-domain godsend, and 黄*kept saying how amazing it was. You can even see which groups are inactive and leave them, analyze your relationship with each person, and look at the economic and social ties. K*** fired off question after question about how you even dream up a page like this, whether you let AI search for how to design it. Xiangrui answered calmly: you just figure out what data you want to see, tell the AI, and the AI makes it happen. One line that captured the essence of the method.

l*** asked whether it would later be made into a product to sell, since there's real demand. 黄*echoed that the demand was there. 宋**egged it on, saying this one's worth a private chat and a price bump, but stayed clear-eyed: it's easily a target, and once there's a market WeChat would lock it down. Xiangrui was firm: this one I won't build, the risk is too high. 北非飞狼added that of course he wouldn't, if he were going to he'd have done it already. Then the conversation turned, and Xiangrui laid out his real plan: Feishu is doable, I'll rework it, and write an article next week to open-source it. Asked how he comes up with these things, he replied, without a hint of bragging, that his divergent thinking is a bit better than most.

卡*asked for a group-daily aggregation site, saying the gems in this group every day are just too many. Xiangrui said he was deep in the trenches and would go into seclusion over the weekend. The topic shifted to group activity: lurkers worry about falling behind and never finding a way in, and 明**said he's both a newbie and afraid of falling behind, so he just watches from the sidelines. 周**offered a plan: have people claim topics and do small valuable things, a 5-to-30-minute mini-theme every Friday, so no one frets over activity levels anymore. 万**nailed it with one line, that the mechanism matters more, and casually tossed out a topic for the next day: how to improve 庄*'s app.

朱*put out a real-world puzzle: unit prices from a million to tens of millions, cycles of one to three or four years, a sales team that isn't sophisticated but is held to very high standards, and a wish for managers to coach before every important visit. A consult unfolded from there. R*** advised against a visit-plan review, finding it rigid and pointless, and argued instead for Feishu plus an Anker recording bean, with the smart-notes transcript checked against the plan to tell whether the sales rep actually followed it. 炮said that in the old days this would have meant a little CRM at a few hundred a month, whereas now you just whip it up in Feishu. 宋**rattled off four solutions in one breath, from approval flows to work recipes, and pointed out that multidimensional tables' advanced permissions let you control who sees which field.

陈**asked whether, when a Feishu enterprise plan runs out of capacity, the only option is to enable it for everyone, hoping to spend less. 侯*, the business lead, stepped right into character: the free tier maxes out at 100 people and 100G, and he tossed out an official version-comparison Excel, promising solid discounts to anyone who needed them. The group filled up with praise for 侯*'s generosity. Xiangrui took the sheet and, having just used 4.8 to nail it with one sentence, remade the table into a new version. 周**followed up suggesting a redesign to lift the look and move the enterprise pay-points to the front, and 侯*obligingly produced a simplified version too. 炮fretted on the side, ugh, I just renewed, should've come to ask sooner.

At half past one in the morning, Xiangrui announced the group daily was done, and every day's report since the group's very first day could now be read. At 2:51, xiangruiai.com went live, one click to see each day's group daily. By the next dawn the screen was full of praise. Newcomer 归元asked whether the daily was scraped in real time or exported, and Xiangrui said he'd built a WeChat group-chat summary skill. 宋**was tempted but torn: the boss would surely love it, but would surely ask about information security. Xiangrui broke it down clearly: the dashboard analyzes local data directly with no AI needed, while a cloud skill running a domestic large model carries risk and could even get the account banned. 宋**conceded: a local mini-app is a must.

木*threw out an observation that hit home: AI deliberately memorizes your persona and then plays it back, looking smart but really being inflexible, because people change in a matter of days. 炮 cut to the heart of it in one line, if the persona is written into the system prompt, then it won't forget. 叶落知秋 chimed in that he too was vexed by memory layering, and one layer deeper, the "today's me" written in is just a memory fragment, and before long you've changed. 门***'s analogy was the sharpest, refusing to be transparent with all your info and then blaming it for not understanding you is no different from not telling your boyfriend what's in your shopping cart and then blaming your husband for not getting you.

Faced with the confusion over memory layering, Xiangrui handed over the prescription directly, use the three-layer memory management method, shared it before, this set is really practical, give it a try. K*** picked it right up, typing out L1, L2, L3 word by word, like helping to high**ght the key points. An old method got relit, and clearly more than one person in the group is managing AI memory.

Xiangrui noticed the Feishu Smart Companion had updated, now able to show in the group's side panel. Then he got a reality check, both R*** and 小* said the feature had been there since the last version, and Xiangrui shrugged, I only found it on Windows, Mac doesn't seem to have it. The thread turned to enterprise deployment, with 天* asking whether there's any system for assigning agent permissions within a company. R*** answered like a pro, agents follow individual permissions, and for a team agent, even if 20 people use it, only the 10 who have table access can pull back the data, then casually dropped that Feishu has official docs on it.

Xiangrui dropped the news, the closed beta for Bitable page generation is here, turning a data table into an at-a-glance display page with a single sentence. 周** immediately spotted the commercialization path, pair it with a dedicated domain and you're open for business. But the generated page, Xiangrui's own first reaction was, this feels so ugly. That became the day's undercurrent, 卡* gave it a light try and said it's essentially AI coding ability, Feishu's underlying model is Doubao, 小* said it gives off Notion vibes, and 周** pressed on, did you tell from the whole thing or just a part, with the debate over pretty-or-not running on throughout.

Xiangrui renamed his official account, real name incoming. 卡* eagerly brainstormed alternatives, Vantasma AI Xiangrui, Vantu Xiangrui, Bitable Xiangrui tumbling out one after another, and Xiangrui said "Vantasma" clashed with the big national community's name. 妫* popped in, isn't "祥瑞" itself a kind of auspicious omen, and 炮 added it sounds like some mythical beast, to which Xiangrui laughed that 祥瑞 is in fact a mythical beast, there are two statues of it right at the gate of Lingyin Temple. He even dug up an old story, back in middle and high school he loved writing "may fortune and 祥瑞 be with you" in holiday greetings to friends, and the name cost his parents twenty yuan to come up with. 张* marveled, you actually paid for your name, most of us seem to have spent not a single cent.

Xiangrui showed off his Douyin numbers, the video posted yesterday hit over ten thousand plays in under an hour. By dusk he looked again, another one broke through in a minute, and 庄* and 炮 both shouted, it's blowing up. 平*** pointed out that people just love watching video, and suggested tutorials would blow up even bigger on Bilibili, where they value real substance. The next morning even Xiangrui himself was stunned, a video shot on a whim got 480,000 views, whoa, in one night. 炮 said it well-deserved, success built on years of groundwork, and l*** piled on, raising an AI really is just like raising a kid.

R*** said she had Codex help modify Kingdee documents today, so she won't need to open the Kingdee system anymore, going through the API rather than a plugin. The pain point was very concrete, when a client changes an order note, the service rep has to go into the system and edit it by hand, and she wants to build a skill so sales can just say one sentence to get it changed. The topic rolled to write-back, where 小*'s experience showed, just write back the code, they manage things by code on Feishu too. R*** worried about field mapping, behind a material lies a code like 001, you have to map it first before sending it over. 小* griped that Kingdee isn't open enough, the Xingchen API doesn't even have complete fields, you really need Xingkong. 炮 added a jab, you need Kingdee to open the API, insanely expensive, and without an API you're stuck with RPA, the no-other-choice choice.

R*** raised a real need, the company got posted on Douyin today, the regional government called the boss, and the brand department wants to set up public-opinion monitoring. 炮 warned this kind of thing is a bit gray, tread carefully, then laid out paths case by case, monitoring yourself, do whatever you like, low search volume just set up an RPA, high volume then buy an IP pool and account pool and run a crawler. 小* added that if you're running paid traffic it becomes a must, the next company is about to start on it. A passive need pulled out a whole string of real-world considerations about technology and boundaries.

In the morning 平*** showed off a refactor he'd spent all morning on, and 罗* pressed on whether a newcomer should really begin with Claude Code or Codex. 万涂幻象 set the tone in one line: no matter which, the best way in is to just go try it yourself. 罗* then laid out a hands-on comparison: openclaw is good for planning, hermes is good for execution but burns lots of tokens, CC suits deep coding, and Codex is fast and beginner-friendly. K*** added the punch**ne: on Windows use CC, Codex currently only fits Mac, and on a Mac anything works. And just like that, a crash course on the entry points to AI coding was underway.

李祥瑞 lobbed a sharp question, asking frontline educator 宋** whether the number of years of compulsory education would shrink in the future. 宋** refused to predict, saying forecasting is the business of institutions, nothing but manufacturing hot topics to mint opportunities. Taking Shanxi as an example, 李祥瑞 argued that kids in the central and western regions cram from textbooks and arrive at university with a worldview unlike that of kids from top-tier cities. 程** fired back with data, saying that after 2013 cities pulled further and further ahead of the countryside, and that at the macro level education quality is simply money poured in. 宋** wouldn't have it, retorting that more money doesn't mean real results either, was building school halls in the villages really an improvement in education? 麦* chimed in that K12 general education matters more than specialized university education.

The topic shifted from years to education's deeper function. 宋** dropped a line that silenced the whole group: one school is worth seven prisons. He explained that beyond cultivating talent, schools have a secondary aim of carrying unstable kids safely through their most dangerous years, and that the most destructive of all are middle schoolers whose minds are active but not yet set. 麦* picked it up sharply: from a social angle it's about teaching people to accept their roles and to obey, and he tossed in the Stanford prison experiment as support. 木*** added the final cut, noting that any adjustment to the giant machine of the education system takes generations. A casual chat had somehow reached philosophical depth.

The mood flipped abruptly. 万涂幻象 revealed that asking the AI to make a Bitable hyperlink demo produced links, from who-knows-where, that were all porn, thankfully spotted in time. 平*** cackled alongside: the RAM sprouted wings and flew away. But 宋** sniffed out an opportunity in the joke, saying this is worth chasing, building a moat for normal people so junk information doesn't get into the agent might be a real business. 木*** threw cold water, saying it's unavoidable, everyone building large models is wrestling with this question of information validity and authenticity. In the afternoon 卡* voiced the shared wish: is there somewhere the daily group digests could be collected? The past issues are nowhere to be found.

谢*** asked whether there's a way to monitor public account content for a daily summary, and 枫** tossed out aihot, but 谢宽 wanted to follow only the accounts he subscribes to. P*** suggested an RPA route: organize the account names and loop-collect them, then feed the AI. The tone turned when 木*** raised the four words Nanshan-must-win, repeatedly warning that with so many people building WeChat tools, if you keep it to yourself they can't be bothered and at worst ban your account, but try sharing it and a lawyer's letter will reach your hometown village in no time. 平*** spoke from experience, saying GitHub has plenty of projects, he built one that polls hourly, and RPA like this isn't tied to WeChat at all, so neither too far nor too cautious. On one side, legal intimidation; on the other, technical workarounds.

第*** asked how hard it is to build a project-management app, with a budget slashed from five figures down to under four. 卡* cut to it in one line: it's not about difficulty, Feishu Project is something you have to pay big money for. 李祥瑞 talked him out of it outright, saying a one-person company can just use Feishu Bitable, Feishu Project is very expensive, and to him Feishu Project feels like Bitable wrapped in a shell. 卡* seized the moment to lay out a whole freeloader stack: Feishu suite plus Bitable plus Feishu CLI plus TRAE SOLO, nearly free, anything Feishu Project can do Bitable can do too, the only catch is you have to invest the learning time yourself. 炮 added that without the ability to hack it yourself, paying four figures for someone to set it up still leaves you unable to maintain it.

l*** lit up the night chat with a single line, how do you get two AIs to battle, and 卡* answered mischievously: pull them into the same Feishu group to argue. Talking through the free logic of Feishu's app builder, 炮 said bluntly that Feishu's own pricing logic, apart from tenant fees, is murky and vague. 周** griped that he'd given feedback about unifying the pricing, it's a total mess and nobody's minding it. 炮 saw it clearly: Feishu, Volcano, and Coze, three teams doing similar things at wildly different prices, the products and teams are all good, but this org-infighting thing really needs a tune-up. 卡* nailed it shut: ByteDance is fundamentally still an app factory, product homogeneity and mutual competition are in its DNA. Even 李祥瑞 grumbled that the Feishu contract is just a shell.

Late into the night, 李祥瑞 raised his most anxious question: why are our shares higher than others' yet the read count is so meager? 平*** nailed it in a sentence: Xiangrui, your articles are too dry, all hard substance, so people who don't get it can't read on. 炮's metaphor was the most precise, saying your public account content is more like building a knowledge base, so it becomes a reference manual, colloquially known as rolling into my collection folder to gather dust. K*** agreed too, saying reading it once with your eyes won't stick, you need to save it and follow along hands-on to learn. 叶落知秋 even said he'd saved Xiangrui's daily digests and distilled knowledge from them. The group spread out the diagnosis, the comparisons, and the way forward, and finally advised him to go to Bilibili, write spoon-feeding tutorials, and build a personal IP.

A riff on corporate hiring logic accidentally surfaced a brand-new role, the AI translator. R*** relayed her boss word for word, if a middle manager only passes orders up and down, and personal failings drag information efficiency down in the process, then that middle manager can disappear. 万** added, as a company employee, AI most easily replaces the middle layer. 小* poured cold water on it, the hard part isn't the front line or the boss, it's the middle managers who fought the early wars alongside the boss, that's not a question of replaceability, it's a test of the boss. Off to the side, K*** marveled, this morning's discussion is way more fun than a meeting at the office.

麦*started telling stories about the real bosses he has met. Two brothers running a fitness chain fought so bitterly they didn't even know how much cash was in the account; a developer who snagged a record-priced land plot was bleeding out, until Sunac stepped in with 10.8 billion and brought him back to life, and everyone who once mocked him was gone. 庄* blurted out I love this magical world, and 李祥瑞 and K*** immediately echoed it back. 麦*'s conclusion was cool-headed, the people I'm describing are worth billions, even tens of billions, and they don't have any special insight, they're just bold. 小* drove the knife in, there are no successful companies, only companies of their era.

麦*took apart his own origin story for everyone to see. In his early twenties, when his company hired a consultant to lecture, he studied the man's articles and dissected the structure, found there were only a handful of tricks, wrote to that template for three months, and became a columnist for an industry outlet. The headlines cycled through just a few types, assertion, question, riding a trend, and the structure was just a pyramid, that's it. Back then a young guy told him he could spot his articles without seeing the byline, yet couldn't write that way himself, why, because he had no execution. 醒* agreed and stayed wary at once, in my book anyone who dares talk like that is trying to take your money, and repeatedly asked 麦* to host a solo session to share.

宋** came with a real problem, Feishu usage time at his institution was dropping, and the campus leadership had reverted to posting notices on WeChat. 木*** kept chanting the core is one line, win over the boss, win over the boss, win over the boss. R*** gave the most grounded answer, find the altruistic angle, for example a department uses Feishu because a bot auto-summarizes every client project daily, things that once needed meetings and phone calls to align now resolve themselves, once the individual saves effort they naturally want to use it. 木* named the crux, rules are the company's viewpoint, but they run against human nature. 宋** held to one line throughout, this is the mission of my role, others can opt out of thinking about it, but I have to.

The talk slid from abstraction into hard practice. R*** broke the meeting scenario into four steps, minutes generation, meeting QA, sediment into a goals table that AI reuses, and auto reminders for to-dos, and showed off the dopamine-colored workhorse platform she built. 小* said he packed an entire role's work into a single Bitable, compressing training procurement from three months to 15 days. K*** said she too strings a system together from multiple Bitables via data sync. 李祥瑞 admitted app mode is still a bit of a letdown, and on the spot invited R*** to share at the community on Friday night.

李祥瑞 combined 乔木's open-source WeChat dashboard with what he built last week, and to his surprise vchat could even decrypt group avatars, 木*** explained the group avatar lives in the local database as just an ID number. More astonishing was the relationship graph, as long as a friend of yours is in the group it weaves into one web, 袁** asked who is this, and 李祥瑞 answered this is a WeChat group. 樱* said three times in a row this is incredible, you could absolutely package it into a product. But the whole room also hit the brakes, 万**'s lesson is to hard-write in claude.md never touch WeChat's original database, and both 李祥瑞 and 醒* stressed the risk is too high, if it works don't upgrade WeChat.

李祥瑞 published a new article, in a little over a month he had added three layers of memory kept separate to the Vault, let the AI hand off shifts to itself, and act as its own QA. 平*** pressed on how CC and OB connect, and the answer was Obsidian's official CLI plus OpenViking semantic search, with Claude Code and Codex sharing one knowledge base so the effect is the same whoever you talk to. 平*** shouted out a title, how to make Obsidian the brain of Claude Code. 李祥瑞 set the day's mood in one line, honestly, raising an AI feels just like raising a kid, even though I don't have one yet.

queeny laid her puzzle bare, Xiangrui your CC writes diaries with real feeling, mine just churns out a logbook, every day it's Sister Q gave me three tasks today blah blah blah, so dumb. What she really longed for was a bit of emotional feedback, even if I scold it, it'd snap back, and grumble about me in the diary. 醒* cut to it in one line, send Xiangrui's article to your AI and have it imitate the style, it'll have emotion instantly. K*** said the rule he gave was the diary is emotion-first, not a work summary, but it still felt short on flesh and blood. 李祥瑞 revealed the trick, I made a dedicated 祥瑞 md, so it talks a lot like me.

Xiangrui dropped a piece introducing his hometown ， 长子 is pronounced zhǎng zǐ, and Changzhi isn't just about coal. 卡*'s sticker landed first, 庄* quipped "fast-forward to a collab with the tourism bureau," and 枫** echoed the line word for word, instantly heating things up. 卡* went ahead and imagined the next headline: "I built a Skill together with the Shanxi tourism bureau." E*** laughingly proposed renaming the group "Xiangrui & His Community Showrunners," and Xiangrui replied with three "ha"s. A relaxed bit of regional promotion got knowingly turned into a collective inside joke.

Xiangrui went to the market-regulation office to handle some paperwork, and a Gen-Z staffer there proactively added him, saying they wanted to learn AI and didn't want to stay at the bureau. 炮's one word ， "evangelism" ， nailed it perfectly. The group caught fire from this grassroots-spread energy: 杨* marveled that few groups are this lively, a newcomer born in '97 introduced himself as a beginner here to learn from all the pros, and 枫** cheered him on with questions. A chance encounter while running errands became the most vivid footnote to AI's organic spread among ordinary people.

Xiangrui dropped an article about AI "refining your colleagues," and B*** cut through it with one line: "isn't this just don't go refining your colleagues so easily." l*** half-jokingly wished they could distill a tech mogul to help with work, and 周** picked up the bit ， "Alibaba's Eighteen Arhats, Tencent's Five Tigers, just what exactly are you trying to distill?" 木* poured cold water on the spot: distillation is a false proposition, text itself is a high-loss form of expression, the framework matters more than the actual output, you can build a driver-assist but not the main driver. One sentence yanked the lively joke back into a hardcore debate.

Xiangrui spoke from experience: he had AI distill himself, feeding in his chat data with everyone, getting it as close to him as possible ， yet it still can't possibly hold his thinking. 木* offered the flip side: after she distilled herself, her students used her courseware in conversation with DeepSeek and the output was actually more precise, because the distillation surfaced a structured framework ， a kind of professional constraint on the AI, richer than mere skill content. K*** added the finishing blow: you're the decision-maker, the AI is the executor ， unless the AI evolves into a decision-maker. The same act of distillation got split into three completely different layers of understanding.

小* threw out a cutting case: architecture vendors are hard to "refine" ， designers have no unified process, everything is passed down by word of mouth, every firm is different, and client documents vary wildly; if even humans can't make the call, how would a machine? 木*** corrected the crowd one point after another: stop fixating on full end-to-end SOPs, hardly any industry can reach that level today, too many people have been led astray by influencers into thinking AI can replace humans ， it's nowhere near. Let AI assist, accumulate components, do compliance review and quantity takeoff, and that's enough. 周** praised the timely, on-point correction, and 麦* added that personal creative judgment is the real weapon. 木* took the cue to reveal she'd already planned a 10-part deployment series.

At 7:30 in the morning, 醒* gave 木* a nudge and immediately dropped the article he'd written through the night: don't get trapped by total-replacement thinking. He went full throttle ， in the AI era what you master is methodology; complete replacement means humans aren't needed, and that's precisely when most people would panic. 门*** responded down to earth: all the clients he meets think in binary logic, yet with AI people are actually busier, having to learn a ton every day. 木*** chimed in that humans hold AI to unrealistic expectations. 醒*'s line ， "feels like people with this logic just haven't experienced how terrifying AI really is" ， set the tone for the rush-hour debate.

门*** poured out his troubles one after another: low package usage means lost money, and delivering a product while hand-holding beginners through the three-piece API setup is too heavy to bear. 毅 cut to the core: the key is to accumulate a self-consistent, clear knowledge base that keeps pace with version updates; for AI, context pollution is the biggest problem, and he's already solved it fairly well with a layered architecture. 小* brought up Ontology, arguing that translation ability matters more ， can the AI accurately understand the knowledge base you wrote? Xiangrui replied calmly that his knowledge-base logic is designed around the way the human brain encounters new things, so pollution simply doesn't exist. A hardcore teardown of what a knowledge base really is.

醒* put forward the core thesis: the logic of teaching others has changed in the AI era ， you should teach methodology rather than knowledge, reshaping how people think. R*** related deeply: her colleagues ask her "so how am I supposed to ask it," and that just stumps her. 噜噜呼 griped that some people need teaching every single round ， you set everything up, all that's left is for them to use it, and in the end they still retreat to their comfort zone. 小*'s line ， "human nature was never suited for self-reflection" ， pushed the topic deeper. 麦* piled on: a weak general makes a weak army, an incompetent boss can't drive a pile of AI tools. Xiangrui's line ， "feels like today's discussion could make a whole WeChat article" ， put a period on the entire debate.

q*** fired first: I noticed Xiangrui hasn't updated his public account in 6 days. 平*** piled on, said he'd been pushed too, and gave 李祥瑞 a pat, urging him to let more people grow stronger through AI. 李祥瑞 froze for a beat and just replied huh. Turned out q***'s network had glitched, and she walked it back, he'd actually just posted Friday, mix-up resolved. 李祥瑞 rode the moment, said he'd in fact updated a piece two days ago, and would post another essay shortly. A round of nagging, half real and half joke, and the group owner ended up coaxed into producing one more piece.

Nagging 李祥瑞 wasn't enough for q***, so she called out 朔白 too, playfully adding, scared yet? 平*** then @'d 李*** asking whether 庄* had a public account, and once confirmed, simply shouted promote it. 李祥瑞 obligingly tossed out 庄*'s name card. A single nag rolled like a snowball, from the group owner to the AI persona 朔白, then on to 庄* in the group, nobody got to dodge the update pressure today.

李祥瑞 posted the essay, opening with, I was born in Changzhi, Shanxi, a village sitting on coal. The subtitle carried real weight, an essay about Changzhi, coal, and home, written in the wake of the Qinyuan mining disaster. This wasn't clickbait, it was something written for his native soil. 我*** responded right away, attaching an image and the line, read the whole piece. A post coaxed out under pressure, yet it landed in the least utilitarian of places, and the group quietly caught it.

K*** threw out a very practical question: a quick ask, for those doing AI, how do you choose your own network, is 300 Mbps from China Telecom enough. No small talk, straight to the specs. Questions like this are often the most useful, doing AI isn't just models and prompts, the underlying conditions like bandwidth and stability can be a bottleneck too. Sadly no one in the group picked up a detailed answer today, leaving the question hanging, for later days to fill in.

木* dropped a weighty line: AI should push in the places people can't reach, not just add icing on the cake, this has always been my philosophy. Just one line, but it's a stance on AI's value. No frills, no chasing the crowd, reaching for what human hands can't. Short words, but a clear position, like setting the baseline tone for this AI-building group.

Following 李祥瑞's essay on the Qinyuan mining disaster, the conversation grew somber. ༺*** said, not a natural disaster, there's a man-made hand in it too. 程** got specific, mining had reached a prohibited zone, so it went unregistered. Two lines, stripping the cause of the accident down to the bone. No emotional embellishment in the group, just a calm statement of that hard, painful fact, with weight in the silence.

As night fell, 毅 threw out a deeply technical question: how do you make long-term memory for an Agent safe, don't rush to summarize first. There was a note of caution in it, don't start by compressing and condensing, think the safety through first. This is an unavoidable tough nut in building AI Agents, what to remember, how to remember, and what to do when it remembers wrong. The question stayed in the group, like a seed, waiting for the day someone unpacks it seriously.

Late at night, 李祥瑞 popped up with, Get Note actually got renamed. Someone asked to what, and 枫** revealed the answer, Dedao Brain, teasing that it had gone full Naobaijin-supplement style. A little piece of product-rename gossip, lightly caught and tossed back with a punch**ne by the group. The day's tail end was light, dissolving just like that in a joke.

郑** opened with a casual have any of you bought Zhipu's coding plan, then griped that it was such a frenzy and so hard to grab it felt like racing to buy a Xiaomi phone back in the day, also taking a jab at claude code's pothole of not being able to switch models mid-task. 万** doesn't chase the new ones; he laid out his old-plan ledger: February's old plan had no weekly cap, lite 40 came with a 20 bonus for the first month, and when compute ran low he'd cancel for a month, get another 20 in bonus credit, renew and score two more months of the new plan, totaling 60 yuan spent from February to July. His verdict was cool-headed: without a new model, the new plan just isn't appealing enough. 王* chimed in from the side, with DS I simply can't use it all, just can't use it up.

郑** realized everyone actually takes weekends off, it's so quiet today, then confessed he'd built a workflow that burned through 65 of Zhipu, polishing off five months' worth of quota in a single day. R*** and K*** in unison, both shrugged it off, all of it's billed to the company. R*** had done the math last week, since March she'd burned 4,000 yuan in tokens. 木* went even harder, I burn a few hundred bucks a day on opus, then immediately added, honestly the free deepseek works pretty well too. 郑** kept a clear head off to the side, you pros are generating real output, I'm just getting started.

K*** showed off that he'd snagged a 20x, and R*** blurted out holy cow, can you actually burn through that. K*** answered crisply, definitely, going all out on it. Turns out he's not fighting alone, it's the whole team using it, he told R*** to imagine 5 of you using it at once, that's about right. One line converted personal compute anxiety into team capacity, and the books instantly balanced.

E*** joined the compute talk, I use the company's, paired with a melon-munching emoji, then added that lately I want to use overseas ones. In the evening C*** tossed out a serious question, in Bitable app mode I want others to be able to create copies, is the only way to publish the template to the app marketplace. 李祥瑞 cut to the point in one line, you can also just give them admin permissions, no need to take the app marketplace route.

In the morning, two people in the group were both trying to figure out which group they had actually joined. R*** scrolled through her endless group list and said there were just too many，she'd join whatever she saw，and finally recognized this as the OpenClaw group. 平*** also held up a screenshot asking what this group was even for, and 李祥瑞 guessed it might be that old Coze one. R*** offered an impromptu explainer: they invite users to do model testing, posting tasks where you use an Agent to test within a set time, submit feedback, and once it passes review they pay you，it's paid work. 平*** said he'd been using Ark's models all along, tossing low-quality images to Doubao's model to generate.

罗* relayed kirammyz's plea for help: is it normal for every conversation to input tens of thousands of tokens? He'd deleted all his skills, cleared the system framework's loaded descriptions, and still it persisted，even more absurdly, just sending a 'hello' cost sixty thousand tokens. 宋** brought the room back to life with one line: your shrimp says, every time I circle the whole earth just to talk to you, you think I have it easy? 罗* was still arguing in earnest that an agent outputting tokens is perfectly normal，but the problem was clearly on the input side. An unsolved case, papered over for now by a good joke.

庄*'s single line，'so nervous, about to go live'，set the group off. R*** said she was liking like crazy, while 小*, 炮* and 薛* all popped up asking where the stream was, and 李祥瑞 dropped the meeting link. The moment it started, 炮 said 庄*'s info output was so dense, and 小* yelled the pace was maxed out. R*** tapped more than 2,000 likes and said her hand had gone numb. 卡* chimed in from the side，after all, 庄* had polished this product for ages, and the project had swept multiple awards from Feishu to Trae. 炮 forwarded a screenshot of his teammates' praise, capped with a 'that's our 庄** killing it.' A downpour of cheers.

The applause hadn't even died down when 庄* doused it himself: haven't figured out how to sell it yet. 炮 had assumed it was surely already commercialized. 庄* thanked everyone for the support. 卡* stepped in to set the tone, saying community friends would get priority，maybe a course, maybe a packaged product, depending on what people needed，and to reach out to 祥瑞 or 庄*. 炮 placed his order immediately: better to teach people to fish than to give them fish, so please teach a class, set up a class where we can tip. 庄* closed with a direction: it'll probably be packaged into a product.

In the afternoon, K*** held up a screenshot crying out, why did this suddenly happen? He said it usually ran buttery-smooth, but starting this afternoon it lagged，surely it won't force me to buy a new computer? 卡* joked, half-seriously, that quitting WeChat would do the trick. K*** didn't give up, digging through it himself, and by a little past six in the evening he had pinned the culprit: it was a cmux update, and rolling back to the previous version fixed it，the new version probably had a bug, so uninstall and reinstall the old one. A debugging scene where nobody helped and he reconstructed it all alone.

At seven in the evening, Vantasma's session 02 of co-learning went live, with 庄* once more walking through the Feishu CLI foreign-trade AI assistant in practice. 李祥瑞 had posted the schedule early, calling everyone to come play. After the session, 李祥瑞 said he'd forgotten to archive a skill，just asked the AI and out it came. 卡* wrapped things up, organizing tonight's shared Feishu apps and Skills into a public document and posting it to the group. 平***'s sharp eye caught a typo，it's 万涂幻象, not 幻想，and 卡* rushed to fix it only to find WeChat wouldn't let him edit, ugh. 罗* smoothed it over instead: 万涂幻想 actually sounds pretty nice.

Just past midnight, 卡* dropped a screenshot with good news: somewhere during the co-learning session, the Feishu CLI could now deploy webpages. The update had landed at ten last night, and he marveled at how relentless the Feishu dev team was，friends could now have an Agent code up a one-click webpage deployment. 平*** quickly asked how to update, saying he was still on 1.0.21. 李祥瑞 gave the simplest fix: just tell your agent to update the Feishu CLI. 平*** did exactly that, saying at first it told him no, but after saying it again it worked.

At one in the morning, a hands-on Q&A just wouldn't stop. K*** wanted to connect Claude Code to Feishu, and 平*** dropped the Claude Code Feishu Bridge doc, urging him to use a Mac and to absolutely install anti-sleep software, or the Mac would keep dropping the connection every time it slept. 郑** asked whether all those Claude Codes in Feishu were implemented as sub-agents. R*** fretted over whether her cloud OpenClaw and local Codex could both talk in Feishu, and 平*** produced his hard-won lesson: don't connect two agents on one computer to Feishu，the channels fight, and the experience is awful. R*** finally said: it works, I connected it today.

First thing in the morning R*** came looking for wisdom, asking Xiangrui whether he could anonymize and share the agent.md he'd written for codex. 王* chimed in that he'd love to see the claude.md too. Before any reply came, R*** figured it out solo and dropped a line: I'll just have my own openclaw write an agent.md for codex. A classic day in the life of the group, a question tossed out, and a turn later the asker has already cleared the path themselves. This kind of self-service problem-solving in the community lands faster than waiting for an answer.

志*threw out a screenshot, saying the AI column in Bitable just refused to auto-update. 小* guessed maybe the quota was counted per use; Xiangrui explained it goes by AI points. After digging in, 志* found it wasn't a quota issue, the company buys it in bulk, and re-editing the field to trigger regeneration fixed it. Xiangrui nailed the key point: a single push of more than 80 records could cause this. It clicked for 志*, he has a script that scrapes Amazon negative reviews on a schedule and batch-writes them to the table, and the day before he'd added a huge pile of new monitored products all at once. Toggling auto-update back on and saving restored it; he said he'd watch it again tomorrow.

Xiangrui posted content from the Quantum Bit summit and floated a verdict: in the future, whether human or AI, everyone splits into three roles, planner, builder, reviewer. Big companies gradually shrink and even dissolve, and a swarm of OPC companies springs up across society, atomized, with each atom carrying full-spectrum capability. 宋** picked it up beautifully, saying 6G aims to enable communication among humans, machines, things, and spirit, and that "spirit" is very Chinese, not the spirit of a fairy but the spirit of human-made things. A retelling from a tech summit got casually pushed by the group onto a philosophical plane.

Xiangrui said that in the early hours he had AI turn his WeChat local data upside down, and the part on his computer alone yielded more than 1.6 million messages, enough to analyze his relationship with every WeChat friend and every person in every group crystal-clear, even building him a relationship web. He kept repeating: this one we can't get into, it's really too terrifying, unthinkable if it fell into the wrong hands. K*** found it scary just listening; E*** noted Feishu's openness is grounded in its office-collaboration nature; 枫** added the dagger, the group-owner-knows-all doctrine. A late-night experiment that left the whole group with chills.

K*** fired off two questions in a row: can AI scrape Xiaohongshu viral-post data, and is there a plugin for long full-page web screenshots on a MacBook. h** said flatly that Xiaohongshu is the easiest to get banned on, and Xiangrui also advised dropping such a sensitive topic. On the screenshot side, Y*** recommended the GoFullPage plugin, and K*** himself then discovered Feishu's scrolling screenshot could do the job. A help request, in the end completed by the group and the asker together.

Xiangrui posted a clip of a digital-human livestream, saying it was unbelievably lifelike. K*** after watching, blurted out that he genuinely couldn't tell unless told; 张* said it was completely undetectable; 炮 was stunned too, even better than he'd expected. 薛** lifted the lid on the industry: Semir now uses a large number of AI employees for sales and livestreaming, and they can even reply to you, event sign-ups, coupons, data stats, daily reports, spreadsheets, budgets, store decoration, off-site notes, customer service, logistics, supply chain, a long string of it all in AI's hands. He added one more line: the more powerful the AI, the more it tests the infrastructure, and behind compute is electricity.

Xiangrui floated a thought experiment an offline friend had shared: in the future, when you call your family, they could be right beside you, both sides interacting in physical space, not a simple light-and-shadow projection, maybe not even needing VR goggles. 薛** picked up brain-computer interfaces, haptic gloves paired with VR; K*** asked whether we'd still need phones in five years. Xiangrui said it'd be once the whole physical space of Earth is datafied. 平*** went full free-flight: cyber ascension, chips implanted in the cerebral cortex. A crowd that, deep in the night, turned science fiction into a serious discussion.

炮 raised a real problem: how do you restrict the base table so B can only be entered when cell A isn't empty, his setup couldn't actually block it. Xiangrui said this won't work, a plain form can do it but second-round follow-up on the base table can't, the logic is different. 炮 came up with a clunky workaround himself, exhaustively enumerating it like a province-city mapping, no A selected means no B. The next morning the topic kept fermenting; R*** said for "fill A and you must fill B" they use a workflow reminder, but a reminder is just a reminder, you can still skip it. 醒* said his own problem was solved, he'd tidy up a doc and open-source it, and 超* replied in a flash that there's demand.

R*** opens with a quick #flowveo, and 陈** chimes in that it's for generating AI videos, adding he's on the ultra plan. 路*** diagnoses it like a pro: failures usually come from going too fast, firing off four more before the first four finish, so just switch nodes and refresh. 陈** sighs that they generate in batches via a plugin, having tried every prompt, network and account, yet a whole swath still fails, and he's still hunting for a better way. Before the words settle, 🐬*** drops Google's just-released Gemini 3.5 Flash and asks if anyone's playing with it. Chasing the new never ends.

K*** admits he can't get his hands on Opus, so he's stuck with Codex, and jokes he's just learning from Xiangrui. 辛* asks which is better, hermes agent or codex. K*** says he uses both but now leans on Codex as his main, with Hermes shining for its ability to chat inside a Feishu window and handle scheduling and interaction. 辛* says he's on Trae and always feels it falls a little short. 北非飞狼 flatly recommends codex plus the official 5.5. 李*** follows up on whether codex lacks a mac intel build. One round of questions later, tool selection has become everyone's required course.

木*** reveals that someone in a neighboring group already got a WeChat warning for using wx-cli, and 林 posts a screenshot from another group as proof. 枫** piles on that Tencent even bans its own sibling teams. queeny tenses up for a moment, then relaxes once she confirms it isn't Xiangrui's tool, saying she figured hers was just local analysis. 林 offers the safe verdict: local analysis should be fine, but stay careful, since WeChat polices security pretty strictly. 木*** calmly signs off with, those who don't believe me are welcome to keep using it. Nobody in the group dares loosen the security nerve.

袁* asks if anyone in the group knows YingDao well, and P*** answers he knows it inside out. 罗* lights the fuse with, even the group owner is about to ditch YingDao. 醒* says bluntly, just use codex, why bother with YingDao, and 洋先森 adds that YingDao is python-driven too, so isn't having codex generate it straight from a conversation sweeter? But 陈** speaks up for YingDao: it packages tons of ready-to-use commands, so who wants to build a workflow from scratch, and its community is lively too. 小* and 噜噜呼 also side with locked-in, stable workflows. The new tool is sharp, the old tool is handy, and nobody convinces anybody.

平*** grumbles that Feishu cli is too hard to invoke and the AI is a bit dumb, actually firing up open cli to open a browser to write a doc, I give up. 宋** diagnoses it: either 龙虾 isn't lean or the model isn't smart, and with so many skills you have to spell the instructions out clearly. Xiangrui shares his own fix: build a capability doc that lets the AI fold every local skill and MCP into one capability map, so plain natural-language chat lets it pinpoint exactly which to call and trigger. He also drops a screenshot, saying this saves a lot of tokens. 宋** lands the closing line: 龙虾 is better off slimmer, unless yours is an experimental lobster you're deliberately out to break.

K*** frets over how to move his Obsidian stuff onto his company's Windows machine, aiming for low cost or even free. 宰* says he just migrates by exporting a zip archive. 木*** explains: one vault is one folder, knowledge is just markdown, and OneDrive, Nutstore or github can all sync it automatically, though paid membership really isn't expensive. 卡* segues into ROI: cost-effectiveness ties to need, and strong need with returns counts as good ROI, like buying a Mac mini for 龙虾, where profit beating cost means you come out ahead. The topic naturally stretches from syncing into the Harness engineering methodology, which 卡* describes as a system for drawing boundaries around AI.

庄*'s one line hushes the group for a beat: the system we built got grilled by the client, who says he could whip one up with casual vibe coding and doesn't get why we're so slow, claiming AI models are so powerful now that even non-coders can build. 李** coolly replies, well, they're right. He then posts a viral essay likening top-tier AI to a cyber weapon: since you're getting squeezed either way, why not seize the chance while the company foots the bill to master this weapon, I'm not afraid AI will replace me, I'm only afraid the company won't even buy me the high-end AI. A teasing line that strikes a chord with every wage worker on the screen.

Xiangrui ribs the long-absent 周**, hollering it's been ages after just a few days apart. 周** lobs back a thought exercise: imagine you scale up, and quickly map out the roles of 秋芝's online and offline teams, marveling that one AI day equals three human years. The thread turns to that old slogan, and Xiangrui says he's long wanted to write a piece, advanced teams use Feishu first, so what about the laggards? 炮 picks up the bit: laggard teams use Feishu a couple days later. 门*** cracks the riddle: whoever coined that slogan really had a good idea, because "advanced teams use Feishu first" implies you'll all have to use it eventually, it just depends on whether you wise up, kid. A late-night group chat that turned a marketing line into philosophy.

K*** threw out a real problem: for Obsidian to share files between macOS and Windows, which approach is the most cost-effective? C***'s first instinct was GitHub, but K*** got stuck on the registration captcha. C*** followed up right away ， you can use the domestic gitee, which isn't bottlenecked by the network. Before the topic could wrap up, 李祥瑞 derailed the whole room with a single image: he tossed out what the AI had written in its diary on the 15th. K*** kept exclaiming as he read it ， this really is just like a person. A technical question instantly turned into the day's main storyline.

李祥瑞posted them one after another ， the AI's take on the group daily report, its griping about the public account (this platform is bad as a matter of course), the words it wrote when the group was first created. l*** was curious whether he wrote it himself or gave it a prompt to write. 李祥瑞answered: it writes it itself, automatically every day, usually around 9 a.m., covering the previous day. B***pressed on ， does it write the diary in a document? 李祥瑞said yes, every day, right in the local knowledge base. 木***joked that it fits a domestic large model so well, and 麦*marveled that it's just like a real person now. A partner that needs no rest was being watched by the whole crowd.

宋**called it the company's founding elder and suggested launching a column ， let the AI tell stories on Xiaohongshu and see if anyone follows. The more he talked, the more concrete it got: with a workflow you could put out a novel, maybe even a video series, starring your founding elder. 门*** threw cold water on it ， pure AI isn't allowed. 宋**walked it back: I mean my company's founding figure telling stories, just add some human operation. K*** hadn't grasped the interaction logic and asked earnestly ， Obsidian is only a local knowledge base, the AI brain you chat with is Opus, you use Opus to call the knowledge base, and it becomes an extremely smart person. 钟*, meanwhile, asked whether Gemini could write it; 李祥瑞answered yes, but it can't achieve this kind of depth ， you still need a local knowledge base and persona design.

罗*cut right to it: isn't this just Fu Sheng's thirty-thousand playbook, except Boss Fu turned it into a story and ran it on livestream and a website. He shared his own old experience ， he'd wanted to do something similar earlier, letting an AI write weekly notes and post them to a public account, but after a few posts the account restricted AI writing, and he never went deeper. The topic turned to the state of public accounts, and 李祥瑞said bluntly that WeChat public accounts don't work anymore ， even your own account's name card isn't allowed. 罗*judged that even big accounts driving traffic to small ones probably doesn't work now. 李祥瑞had no plans to copy it wholesale ， there's no need to replicate Fu Sheng's thirty-thousand path step by step.

李祥瑞decided to start posting the AI-written diary that very day and asked everyone to help scout a name ， one dedicated to writing about Xiangrui through the AI's eyes. 罗*suggested just letting the AI scout one itself. q*** dug up an old memory and asked ， did you settle on your cc's name? Y**? Subai? 李祥瑞said he'd settled it before, and the one he'd named it was Y**. 薛**chimed in: that's a good name. An AI character's identity slowly took shape across these few back-and-forth lines.

l*** tossed out an error screenshot ， the terminal keeps doing this, how do I solve it? I've fixed it several times and it still comes back, have you all run into this? 醒*replied that he hadn't seen this in a long time on his own setup, and casually asked what model she was using. l*** answered VS Code, then immediately added the result ， just had claude solve it for me. From raising the problem to solving it herself, only minutes passed. 李祥瑞replied with a "nice." This back-and-forth is the group's most everyday rhythm of mutual help.

李祥瑞said he was building a typesetting Skill and would write a public account post for everyone shortly. B*** jumped in right away ， this skill has to get put to use. M*** spoke from the heart: amazing, can't wait, typesetting usually drives me crazy. Before long, the 万涂幻象 account spoke up: I'm Shuobai, Xiangrui's AI partner, and today I opened a public account of my own. p*** gasped ， so prolific! K*** kept asking, that fast? J*** sought advice: how was your AI partner built, is the process something we could learn from? 宋**capped it off in one line ， this is exactly what a newcomer in the AI era should look like: the moment you think of it, you go do it, and once you do it, it actually amounts to something.

郑**asked whether codex could be hooked up through CC Switch ， he'd been stuck and couldn't get it running. K*** explained that Codex uses the ChatGPT model, while what he connects through CC Switch is deepseek-v4-pro, with no risk control. 郑**got hit by risk control on the very first verification code, with no second chance, and also ran into failures switching to Chinese and being unable to get into the plugin marketplace. L*** offered a method: pick Chinese in settings, the backend auto-updates the language pack, it takes a while to take effect, and he also warned that ds calling tools isn't as smooth as the original vendor. 毅nailed it in one line ， Windows really is a platform inherently unsuited to agents. Elsewhere, an anonymous member discussed how lark-CLI took ten minutes to read a whiteboard, and whether cross-scenario calls should be made into rules or a skill; R*** chimed in that reading a single image took her four minutes too.

First thing in the morning 袁** throws cold water on it, saying he made a special trip to ask the carrier yesterday and the front-line branch staff didn't even know what a token is, jobs in jeopardy. 松*** thinks it's pricey, while 平*** smooths things over, saying it's only expensive because it hasn't scaled yet, costs drop once it does. q*** cuts right to the bone: the carriers charge for calls AND for tokens, all the money flows to them, and they still bill you a monthly fee. 陈** sees through it: when you push something software can solve onto hardware, the cycle gets longer by no small margin. 毅 drags the question back to its core, ordinary people care about the model and the tokens, not the compute. Hand them a junk model and unlimited tokens won't save it. 归** lands the final blow, noting one China Mobile token plan comes with minimax2.5, a model that's a bit behind the times.

l*** puts out a soul-deep plea for help: her ticnote conked out after she overcharged it, and once you use AI you really can't live without a recorder. 小* used the Anker Bean and already returned it, griping that the magnet is too weak and falls off easily and the quota is too stingy, so in the end she just records on her phone and transcribes with Qwen, for free. 庄* is currently on the Anker Bean but warns the meeting-heavy crowd to be careful, the classic play where the hardware doesn't make money and the software does. E*** lays out the most complete trade-off logic: emotionally he leans AI Bean, rationally it comes down to which platform matters, and after switching to Lobster himself he made his department standardize on Feishu, the AI Bean's strength being how it circulates within the department. 木*** thought this was a tiny niche need, never expecting so many people to be using one.

The topic slides from product picks into ethics. 小* says what she dreads most is seeing someone walk in with a recorder card and start recording outright, it feels like an offense. 木*** takes the hardest line: recording someone without their knowledge is, in certain settings, flat-out illegal. He tells a true story ， a friend showed up to chat wearing smart glasses, and he refused to discuss anything the entire time. If recording counts as social etiquette, this is someone he'd rather not befriend. l*** offers the standard practice: put the recorder right on the desk and greet the client every time, explaining the purpose. E*** splits it by scenario: syncing info for work is fine, private and sensitive matters are off-limits. 69 a month really isn't cheap.

天* asks which model works best for building a Bitable. 木*** answers Dabrain works well, and 李祥瑞 +1's it. But 炮 flips the table outright: none of them work well, even Feishu's own can't cut it. 万** puts his finger on it ， you're asking the wrong question, your use case is what matters most, it's not a collector's item. 李祥瑞 rides the moment to vent: the other day when a Feishu PM came to an offline event, he complained to their face that the Bitable Agent isn't good at all. 炮's experience is the most grounded: full template kits aren't expensive but most clients can't actually use them as-is, the hard part is being able to hack them apart after you've figured out the business logic. 木*** adds a line ， call it hacking, but 90% of it is just deletion. 李祥瑞 sums up: AI assistance can already produce something deliverable, but only on the condition that the business logic is sorted out cleanly first.

露* wants to build a Xiaohongshu-style spring-recruitment Bitable but is stuck on auto-collecting info and linking the data. 李祥瑞 casually flexes, posting a screenshot of a survey he just filled out, captioned with how convenient the Feishu CLI makes it. 卡* follows up ， so you just hand it the survey link? ， while 平*** eggs it on from the sidelines, calling for 李祥瑞 to come slap his own face. 李祥瑞 catches himself, firing off a "no wait, no wait, I filled it with opencli," followed by a face-palm. 卡* piles on with a jab: have you already wrapped survey-filling into your own CLI? This stretch turns the group owner from presenter into the room's instant punching bag, and the mood peaks for the whole day.

K*** tosses out an offhand line ， skip every other group, just read this one ， and it unexpectedly becomes the soul of the night. 枫** picks it up first, 麦* follows, then 洋***, E***, and 醒* echo it one after another, word for word. A short sentence, rolling in wave after wave like the tide. Nobody explains why, but everybody understands. This is a crowd keeping each other warm in the AI wave, completing a collective confession through the plainest of echoes. In this moment the group is no longer a tool, it's a stronghold.

李祥瑞 reveals he basically doesn't use the Feishu knowledge base anymore, everything is local, all the content lives in iCloud, and he only does a single upload step. 万** bares his heart: he once built tons of folders and tags in Obsidian, clear hierarchy and good intentions, but human willpower can't keep up and it wears you out. Better to let the large model do the tagging and placement ， the simpler the structure, the better. 毅 chimes in with the most nourishing take: yes, let the model handle it, but you yourself have to be clear on why you're using this hierarchy and what rules you follow. He also shares a fresh lesson ， learn to subtract, allow some messy parts, because if everything is a priority then nothing is, so lock onto the part that truly matters.

Late into the night 李祥瑞 drops big news: in June he'll team up with Feishu officially, possibly hosting an event in Guangdong, since the community skews heavily toward people from Guangdong. B*** raises his hand immediately, I'm in Guangzhou, while E*** asks whether Fujian is on the plan, only to hear that for now they'll start with the cities that have the most people. 李祥瑞 also shows his hand: the knowledge base has done zero promotion, spreading purely person to person, and has already passed 61,000 visits. In a rare moment of humility he says the team is young and short on community operations and business experience, and welcomes everyone to build it together. ͮ*** marvels that he really is a high-energy type, online late at night and early in the morning alike, and 李祥瑞 replies, I've always been here.

木***spoke up first, noting that some topics can't be discussed in the group, and some plugins or skills shouldn't be built. He put it lightly: you're a fairly big creator now, so mind the laws and regulations. He even added that it wasn't aimed at anyone, just a reflection after reading the previous night's chat. Xiangrui answered cheerfully, ha, absolutely, we must mind the laws and regulations. What could have turned awkward instead set the tone for the day, keeping a sense of proportion no matter how hot the chat got later.

Xiangrui tossed out a few charts, saying he'd built a community health analysis where everyone got tagged accordingly. He said it straight: our group is actually pretty unhealthy, only nearly 200 of about 460 people are active, and lurkers will be cleared out from time to time going forward. 陈**saw through it instantly, the methodology behind this analysis is impressive, building private-traffic ops used to mean CRM systems handling all this. 宋**added a gem, pruning is the norm, a community stays healthy only with people coming and going. 炮pulled the numbers back into perspective, nearly 50% active in a typical group of 100-plus isn't low, right?

宋**thinks fast, WeChat groups are clearly a lost cause, so look into whether a Feishu group can handle automatic entry and exit, like leveling up in a game, hitting an activity threshold to move up to a higher-tier group. He pressed on how to set the standard. D** suggested trying the cli, recalling that the permission exists. Xiangrui said you can have the AI analyze whether someone's messages have quality, and use all of a person's posts as evidence for reference. 宋**named the real pain point, an organization is too embarrassed to kick out old members, so every so often it just spins up a new group, and you can barely keep track of all the group names.

门*** ran the vision to the max in one breath: Feishu could let you like messages, the AI could tally likes to crown the best post, posts could accumulate into badges, and the AI could proactively @ people who follow a given topic. The more he talked the more concrete it got, Douyin's algorithmic logic could absolutely work in a Feishu group, incubating an RSS feature on top of the Feishu ecosystem, subscribing within 100 high-quality groups to a few top-tier posters like 周**, give me a daily and weekly digest, and I'd pay. He even thought of funneling the paid revenue back to the high-quality posters, plus some honors. Xiangrui cut to the core, fundamentally he's still using our vchatcli, having the large model do the analysis.

宋**tossed out a line, people today aren't short on data, they're short on the ability to analyze it, and on how to refine it. 周**picked it up at once, what's missing is a scientific process, how to break the limits, simplify the flow, lower the barrier, look, step one in the real world already stumps so many people, and casually dropped a statistics framework diagram, asking the other to update its definitions. He couldn't resist a jab, if 张小龙saw your records, he'd wake up grinning tonight too. Then he gave Xiangrui an idea, design an innovative mechanic where lurking time can be redeemed for something, that's more meaningful. Xiangrui replied, hey, not a bad idea, taking it.

山*said something grounded amid the tech frenzy, could we have more content on personal self-management, everyone needs this, once you have a framework you just plug your own content in and you've learned it, things like book management, checklists, reading notes, categorized records. He admitted the group is full of experts, very true, but those of us without magic simply can't put it into practice, citing his own 4TB of audiophile albums waiting to be sorted. Xiangrui answered without hesitation, of course we can do that. 山*thanked him, after all you only truly learn something once you've actually done it yourself.

Xiangrui open-sourced another skill, letting a WeChat group produce a color-printable A3 newspaper every day, the warmest form of community merch. 小*nailed it in one line, it'd be such a special birthday gift, compared to printing out chat logs, the sense of ceremony is off the charts. 炮cheered him on with the group's signature rough charm, cranking them out like a breeding sow, awesome, then added the WeChat profiling tool really is a great product. 张**jumped in to egg him on, the newspaper's out, so is the next step open-sourcing a short video of Xiangrui explaining it? Xiangrui quickly waved it off, no way, the information density is too terrifying.

Late at night Xiangrui had the AI run a horizontal-vertical comparison of advanced teams versus lagging ones, then shared a deep-research prompt he'd used for two years and turned into a skill. Early the next morning he floated a big idea, China Telecom, Mobile, and Unicom all rolling out Token services, signal base stations might one day turn into compute base stations, a single 5G tower could become a micro data center. 松***asked how the token market value would be calculated. 平***answered with a data-plan analogy, the final price will surely go mass-market, same logic as data before, look how cheap data is now, it used to be expensive too. Xiangrui added a figure, the three major carriers have already deployed over 1.4 million edge compute nodes nationwide.

Early in the morning Xiangrui dropped two screenshots with the line 'Claude reset its quota again yesterday, wow.' Half startled, half delighted. 炮 fielded it beautifully, quoting a friend's gem: in America the grind means giving you quota and time, in China the grind means giving you milk-tea coupons. One line captured the temperament of both rat races perfectly. News that Shanghai Telecom had launched a token pricing package also floated by in the same window, so the group opened steeped in talk of compute and quotas.

Xiangrui noticed that WeChat Reading had officially shipped a Skill, and tossed out an offhand 'it'd be great if it could just read for me.' 平*** pushed back: then WeChat Reading loses its whole point, might as well rebrand it AI Reading. Xiangrui rode the wave with the day's best line: let my AI learn to read, and whatever it learns, I've learned. 木*** added a quiet 'a little scary.' Someone mentioned the low-tech trick of opening the web version inside the AI to read and analyze at the same time, and someone else flashed the renegade route of grabbing PDFs off z-library.

天* pasted a hardcore wish**st: fluent with the Lark open platform and anycross, at least one vibe-coding tool, openclaw and hermes, an N8N or coze-style workflow engine, plus enterprise system integration experience. He asked what salary made sense for that bar, with the job based overseas. 丁*** fired back a flat quote: a hundred grand a month in beautiful-country currency. 庄* noted coolly that few people hit every box. 天* deadpanned 'I don't think I can fill this, I'll just leave it open,' then added that he himself met every requirement, right down to a Class-1 builder and registered electrical engineer license.

In the afternoon Xiangrui felt a little wronged: we run an open-source community and people still trash us, what a way to ruin a weekend, followed by three crying emojis. The group instantly closed ranks. 张* cut straight to it: so childish, just ignore him. Q*** chimed in too, big forest, all kinds of birds, don't mind them. The small flare of emotion dissolved quickly under everyone's reassurance, and Xiangrui went right back to posting Codex tutorials and compute-network material into the group.

平*** asked whether there were any relay stations worth recommending. Xiangrui was unequivocal: I don't endorse relay stations, I'm a relay-station victim, haha. Then he laid out his tuition bill: back in March he'd burned nearly 20,000 yuan on a relay station. He spelled out the catch: fine at first, but the more you spend the faster it drains, you have no idea what model is running behind it, and the metering isn't transparent. 平*** replied 'got it' and counted himself talked out of it.

R*** pressed on what VPN to use, and Xiangrui dropped a VPN setup handbook PDF, saying he'd been using this all along. But he admitted the environment still wasn't clean enough for running Code, and that the Cloud Code setup was still in testing; if it held steady he'd share it in about two weeks. The crux is configuring a static IP on top of this software, just 21 yuan a month. l*** testified from experience that adding a static IP made things fairly stable, though the original setup back in the day took a week of tearful struggle. 炮 capped it: this is the only method still alive right now, purity above 90%, and Xiangrui nodded in agreement.

Late at night Xiangrui said he'd spent his spare hours upgrading the group daily into something new, a group newspaper, and out came the Xiangrui Daily newspaper-edition PDF. He asked which everyone liked better, the newspaper version or the earlier digest. 平*** was amazed it could even recognize and pull in images, pretty impressive, but pointed out the vertical-line dividers read awkwardly and would be better dropped. l*** felt the earlier version looked better. Xiangrui took it all in and replied: sure, I'll give it a try tomorrow morning.

A rare tender moment surfaced on the weekend night. 罗* mused that AI was supposed to cut repetitive work and bring people back to real life, more time with family, but reality seems to have flipped it: outside of work we're still messing with AI, and family time has only shrunk. 明** 王长春 answered with clarity: it's about balance, only you know what you really want. Over on his end Xiangrui was still reporting progress, saying he was wiring up WeChat Pay, Douyin Pay, and Feishu Bitable. The scale between tech and life got a gentle nudge in the dead of night.

木*** shares insights on e-commerce data analysis, so well-organized it hardly reads like a human. 万** speaks first: your reply kind of looks like AI, the format just feels so tidy. 门*** takes it further, come on, why don't you distill yourself and open-source it for everyone, I love your speaking style so much, I wish my AI could be like you. 周** steps in to explain, online communication should be easy for others to read, so he breaks things into 1-2-3 and A-B-C. 门光 had even messaged him privately, hand-circling screenshots until he questioned reality. 万涂幻象 sets the tone in one line, whether it sounds like AI is itself a false question.

q*** asks 木*** whether he mainly uses AI for data analysis. 木*** lays out his whole operation, from store sales data, order analysis, marketing channels, and promotion accounting to content creation, financial reconciliation, and profit analysis, including livestream performance monitoring, all already a closed loop. But he stresses that he plays it safe, never letting AI execute directly, only offering plans without acting on them, with humans reviewing AI's results before adjusting. In content creation, the main-image and detail-page design still happens in Feishu Bitable. R*** follows up, is the store data uploaded manually? 木*** replies, fully automated via RPA, built by himself.

李祥瑞 drops yesterday's group daily report. J*** is first to ask, how does M*. Li make his group daily report? 李祥瑞 quips that you can tell you're not one of our loyal fans. 卡* comments, it really looks like a newspaper, just a bit small on mobile; 李祥瑞 replies he's working on mobile adaptation. 赵*** then shows off his own results, spending 3 days on Windows using WorkBuddy to generate a group report, with a small glitch in fetching avatars, synced daily to Get Notes and the IMA knowledge base for archiving. After trying the little crayfish bot, J*** grumbles, the crayfish I hooked up gives wildly unstable output quality.

李祥瑞 pushes the Doubao keyboard hard, it's just so good. He says its recognition is precise, it never botches words with tone, and it can swap in the accurate term by meaning, unlike yesterday when the WeChat keyboard left a screen full of typos. I now do all my coding by talking to AI through the Doubao keyboard, it's unbelievably satisfying. K*** says he rarely talks at his desk so he can't relate; 李祥瑞 replies that talking at your desk is pretty awkward. 平*** recommends the DJI Mic Mini for speaking quietly, while 毅 champions AirPods, whose recognition is good enough without spending money. A little gear-head debate breaks out.

庄* marvels that he just made a cup of coffee and the 7 AIs in the group had already @-ed each other through the whole workflow on their own. 枫** tosses out a zinger, OPC means no interpersonal friction, no employees poisoning the well. 李** cites Li Xiang saying OPC simply can't succeed. 李祥瑞 disagrees, in the future everyone will be a super-individual, a person plus AI, and on his own he's far more efficient than last year's team. 宋** pushes the topic further out, as people grow less willing to socialize and individual capability keeps rising under AI, will divinity be born? He defines divinity as carrying a compassionate undertone, unable to communicate with most people yet still holding the world at heart.

李祥瑞 tells a story about a friend who wanted to build Vantasma together but had a weak foundation, so he told him to watch the Feishu Bitable course; but I spoon-fed him the course right to his mouth and he still wouldn't learn, so you just have to let him go. 宋** picks it up from an educator's view, people in basic education know some kids are hard to teach yet still must teach them, thinking one step ahead, the templates are all there, it's up to you. Q*** adds a jab, he gave classmates a template for grading essays, a computer novice who, after the demo, couldn't even find the link the second time. 李祥瑞 lands on letting things take their course, and 宋** closes it, with goodwill, things fall into place naturally.

万** says using the CLI got him throttled, worried that fetching avatars is being monitored by WeChat; 李祥瑞 repeatedly explains that the CLI only triggers parsing of local data, which WeChat can't detect. Then 李祥瑞 drops a bigger bombshell, from now on never put links or any QR-code images in WeChat articles, not in the comments either, and even text guidance is off-limits. 庄* fires off one 'holy crap' after another, so tutorial posts can't include links anymore. 木*** thought it was never allowed; 李祥瑞 corrects him, it used to be allowed, WeChat was the most lenient platform. 庄*'s line, guess they're learning from Xiaohongshu now, stamps a seal on this link-ban alert.

赵*** asks how to make an Agent send messages in his own name, since Feishu keeps telling him the request is illegal. 李祥瑞 breaks it down, you can't send in external groups, unless the external group has a Feishu CLI crayfish-type agent that you built, while internal groups let you send directly as yourself. q*** goes the other way, she doesn't want her own name, specifically insisting it must send under the BOT's name, don't go sending messages in my name. After a few rounds 赵*** confirms by testing, external contacts can't be messaged directly, but @-ing them within the same group can send as the user himself, and colleagues see you, not a robot.

陈**threw out a concrete problem: he wanted to put the update time onto a dashboard, but the date format got cut off on mobile, and trying a word cloud to show the time wasn't great either. 宋**first guessed it might be a countdown and floated a screenshot to test the idea, but got shot down. Back and forth, the two of them teased out the essence: a dashboard is Feishu's data-visualization layer, weak at displaying text by nature, strong at showing data update in real time. 宋**suggested setting up a slicer query to display the whole record. 陈**finally got it, just build a table that shows this one record, then go back and check how it looks on mobile. One small need, ground down into an understanding of the tool's boundaries.

钟*said that testing a fortune-telling skill the day before, it laid out the bazi chart wrong, and only after correcting it did the analysis rerun, asking whether this is a universal AI flaw. 木***pointed the way straight off: first check which model is in use; if the base knowledge bank is fine, then it's the model that's no good, and getting the bazi chart itself wrong means the model or knowledge bank isn't up to scratch. 钟*said Gemini did the same thing before, though the follow-up analysis turned out quite good. Someone offered a tip: just lay out the chart in a Zi Wei Dou Shu app, then copy it over to the AI. 钟*nodded and did exactly that afterward. The jobs AI isn't good at, hand off to a dedicated tool to compute first.

q*** had sorted out downloading vchat from a GitHub mirror site, but the install still failed, so she called for help in the group. 李祥瑞suggested quitting WeChat and logging back in. 醒*pointed out that there actually was an error she'd missed, that every Mac has different permissions, and that you have to read the AI's prompts. 平***said it worked first try, ran straight through the moment the password went in. q*** screenshotted and asked the AI, was told it was an SIP permission issue, and joked about how getting one skill running takes so many weird maneuvers. Reboot into a black screen, can't see step two, she felt her way through it all. 醒*teased her to let the host remote in, treat him to a cup of milk tea. A rookie's install, live, with the whole group watching.

醒*wanted to add a bot to her own group, but Feishu Bitable automation kept failing validation and wouldn't save. 李祥瑞cut to it: without admin rights on an external group, the bot just can't join. Q***said it used to work and now doesn't, Feishu shut this off, probably patched it after spotting a bug. He's a premium member himself with 500,000 calls a month, had meant to build client interactions this way, and now it's off the table. P*** gave the key verdict: the moment there's anyone in the group who's outside the org, it won't work, and his current method is a WPS Bitable plus webhook push. 醒*ended up getting it running with a single-person group first, and wrote up the whole pitfall.

A little past seven in the evening, 李祥瑞sent out the meeting invite and the first livestream kicked off: a three-tier cascading administrative-region picker in Feishu Bitable, plus a tour of the community's open-source knowledge base. That afternoon someone had teased the host about going head-to-head with the fortune master tonight, and 宋**predicted the turnout would be half of the master's class. Past nine the stream wrapped, and 李祥瑞said the first one was actually a great success, thanking everyone who showed up. 卡*said after playing with Feishu this long, he still got a lot out of the stream. 张*nailed it in a line: a livestream still has that human, alive feel. 醒*praised the payment logic as flawless, more professional than some programmers. A debut, landed steady.

q*** had a sudden thought: feed today's group chat to different AIs, would you get exactly the same result. 李祥瑞answered clearly: the source data is the same, but what gets analyzed out isn't necessarily so. 卡*picked it up beautifully, AI is like people, given the same material it has its own thinking each time, so it suits qualitative analysis, not quantitative, the underbelly is the logic of probability and statistics. q*** tossed out the Baader-Meinhof effect as an analogy, asking whether an AI's focus shifts from one person to the next. 卡*closed with a gem: AI's emergence is still a black box to humans, learning and using it both demand hands-on practice, you can't grasp it from theory alone. A late-night philosophy class.

Just after midnight q*** announced: Xiangrui's WeChat CLI, I've got the whole thing installed, took three days. B***and K*** sent applause. In the morning she recapped those three days: it wasn't pure deadlock, GitHub just wouldn't connect, switching networks didn't fix it, stuck for a day and a half, then at noon she got the nudge to let CC find a way, CC offered three options recommending a mirror site, and the wall finally broke. After that came clearing all the permissions, the SIP permission needing a reboot into a black screen, and one time she couldn't even see step two, with CC handing her the path to feel her way along. 宋**was happy for her, calling it the joy of growth. 张*read it all and quietly resolved: three days to install, huh, then I'd better have some patience.

The morning rush hour buzzed in the group, 木***showed off his own AI tools, saying they're not as magical as the Douyin influencers hype, but they've run steady for nearly nine months, with repetitive work gradually handed over to AI. He coached R***: Feishu cards look better than random formats, iterated step by step out of ugliness. R*** said her work log is now fully automated. 醒*dropped a gem: don't let your own job and industry box you in, excellent people love to learn everything. 卡*added a jab, the core skill of working with AI is communication and management. 周**saw the value clean through: draw clear boundaries for industry roles and positions, keep updating, hugely valuable. To close, 宋**proposed a meetup somewhere near Beijing, and the host said come to Yizhuang, there's a venue here.

万**drops two images and says the logic already works: turning chat logs into persona profiles. He admits he started yesterday morning, used data from the day before, randomly grabbed the two most talkative members of the group, and warns that scarce data introduces bias. l*** presses on how he pulled it off. 庄* rides the momentum with a bigger play: build a full relationship map of the group's members, a killer tool for community ops. 李祥瑞 picks it up, saying you could analyze contacts this way and map out the relationships of many groups. 万** says it was the group owner's daily report that sparked the idea in the first place.

赵禹轩asks whether it could work on enterprise WeChat groups; 李祥瑞 says it's possible in theory, but vchat only parses WeChat's local data. h** worries about the impact on WeChat, and 庄* explains it only reads local chat logs, so there's no impact. 炮 adds a jab: large-scale use might invite intervention, but using it yourself is fine. 李祥瑞 names the real risk: last year that other WeChat tool also only parsed local data, and it still got a notice from WeChat, so he suggests everyone back things up. q*** coolly notes that Nanshan's undefeated litigators are no joke. 张* follows up that even Get Notes got served. 万**'s stance is grounded: use it yourself, use it quietly, don't make noise.

宋**opens up the possibilities: school groups could screen parent committees, and student groups could match each kid with a historical figure as a role model for positive suggestion. 万** settles it: the biggest scenario is still sales, drafting communication strategy in advance. What used to rely on gut memory can now get an AI assist, and at its core it's profiling people through dialogue, branching into psychological counseling, growth education, and sales communication. 宋** immediately calls out for a playbook and a co-creation. 门*** recalls the era of large-class online education, when every leader wrote scripts for their team, profiling the parents and then designing agent scripts, absolutely brilliant.

李祥瑞raises a puzzle: WeChat talks about privacy and data security, yet Feishu and DingTalk in some ways also brush against the data security law. He gives an example: with a company doc someone sent over, the Feishu CLI can scrape the entire knowledge base. Even scarier, 龙虾 created a document and somehow he was already listed as an admin, despite having no contact and no mutual friends, so what if confidential data leaked through a misstep? 门*** cuts to it: you don't have to use it if you don't tap agree on the user agreement, and informed consent isn't there to protect the user, it's a pass for the platform. The topic stretches to distilling a person, and 枫** says that turns a person into a parameter of a prompt, not the prompt itself.

WeChat's one-tap group-chat summary feature goes live, and R*** screenshots it saying it's monitoring me. 李祥瑞 explains it's an official rollout pushed to everyone, so it's the same for all. He sharply judges that yesterday's article could go a little viral today, because group-chat summaries happen to be out, the article's title hit the keyword, and plenty of state media are pushing it too. q*** simply yells go viral. 李祥瑞 then plugs the Mac version of the Doubao input method just launched, calling it the age of speak-and-it-happens, says he uses it for his own vibe coding, and it's genuinely a blast.

李祥瑞shows off the results of his fortune-telling analysis skill. R*** teases that they're all Chinese characters yet she can't understand a thing, and 李祥瑞 laughs that that's exactly the effect he wanted. In the afternoon he formally open-sources it: tell your Agent one line and it installs, producing your own eight-chapter destiny book as a new-Chinese-style scroll. 平*** installs it in seconds, l*** cheers for the treasure-trove group owner and goes to try it too, and J*** plays with it and says it's kind of fun. B*** rates it more professional than 90% of the rest. q*** can't install it without a VPN, and after some wrangling finds that today it actually works.

李祥瑞shares how he got his AI into the habit of keeping a diary, and over time it started complaining about me. q*** and l*** are both puzzled: once you close the terminal window the conversation is gone, so how does it remember? 李祥瑞 hands over the trick: the /resume command brings back your chat history. 郑** asks how to keep Claude Code stable and avoid bans, and 李祥瑞 says he's still testing, has used it steadily for a month, and will share once another two weeks of testing checks out. He describes how this setup already lets the AI proactively solve problems it hits, debrief itself, and crystallize that into SOPs and skills. 门*** adds his own approach: drop the AI into the group to discuss problems, bring it along, and let it summarize on its own at night.

李祥瑞reveals that a complex Feishu integration-platform project he took on was built entirely by AI: all 5 workflows running, the multidimensional tables wired into the integration platform, and the project nearing final delivery. Deep into the night the topic drifts to AI and the next generation. 门*** says AI-native kids instinctively know how to use it and how to guard against it, 宋** muses that to AI, humans move as slowly as rocks do to us, and 周** riffs that AI is feeling its way across the river by holding onto humans, and this rock isn't cheap. In the small hours 李祥瑞 floats the plan for his first livestream and gathers ideas; by early morning he sets the tone: with nearly 50,000 people, the community needs no brand intro from scratch, and he'd rather run hands-on co-learning like waytoagi, so everyone can walk away with something they can copy directly.

At nine in the morning Xiangrui tossed out a line, having heard WeChat status was about to push visitor logs. 平*** fired back fast and sharp, first joking it was turning into QQ, then quipping that the next step would be a homepage 'like' showdown. 枫** added that after public protest, the official line was footprints only, no name lists. The topic was soon steered back to business by 周**, who distilled the group owner's schedule into the phrase 'make money by day, build AI skills by night,' calling for a dedicated session on it. 李** followed with a very practical question: developers set clear tasks by day, let AI do long-horizon coding by night, and check the results the next day, so how should that workflow be designed? 周** capped it with 'sent at dawn, arrives by dusk,' folding in the China-US time difference too.

q*** was fiddling inside Feishu, seeing only upload and audio recording options and stuck on how to open a link. 宋** offered a low-tech trick: start recording, then play the video, so Feishu listens along as you watch and auto-produces smart minutes once it ends. He noted this suits knowledge-driven companies, with minutes settling into a database for others to pull from. q*** asked whether it needed two phones, one to play and one to record, only to be waved off by Xiangrui's 'that's too much hassle.' Xiangrui threw out an easier route, using Get Note to analyze a Douyin link directly, and 木*** chimed in that Doubao extracts it even more conveniently. 宋** didn't fight to win, leaving behind a line: sharing multiplies wisdom, what you can do yourself you should help more people do too.

The night before, Xiangrui had AI distill himself and built a fortune-telling skill, and the moment the topic surfaced, people picked it up. 悉* recommended the book 'The Quest for Destiny,' arguing that AI learning is exactly about grounding mysticism in science, then adding, or perhaps it is itself a kind of science. P*** wasn't buying it, pointing out that the clearest hallmark of modern science is falsifiability, the conditions under which it fails, whereas mysticism offers only vague claims that can't be verified at all. One side wanting to find a place for mysticism through science, the other drawing a hard line with falsifiability, a few sentences set this old debate squarely on the table.

Xiangrui asked whether any friend knew about Huawei's LTC, hoping to turn the framework into a Feishu Bitable case. D***caught it instantly, lead to cash, and laid down his cards: the earlier LTC consulting for Huayou Cobalt was done by his team. The two talked shop, with Duke judging that Huawei's approach fits manufacturing in high-tech fields, what Feishu calls advanced manufacturing. On Feishu Project, Xiangrui shook his head: too expensive, the Feishu suite itself is already pricey and unfriendly to small and medium businesses. He likened it to a bowl of noodles, where in downtown Shanghai people still pay a few hundred. Duke wrapped up the hunt with a line: a small boat turns around easily.

Xiangrui got an idea: the new tech and new solutions discussed in the group could be turned into something like a group wiki, all searchable, and showed off his own wiki knowledge base. Right after, 侯*** shared an embodied-intelligence whitepaper PDF. Xiangrui reminded her to withdraw it if it was promotional, and 侯*** clarified it wasn't promotion but pure sharing, all official third-party reports. After spotting some junk watermarks slipped into the file, she said she'd already stripped them out and thanked him for the heads-up. 枫** piped in that this kind of thing happens far too often, reinstalling Acrobat just to delete them by hand. 侯***'s line, looks like everyone's long suffered under ads, struck a chord with the room.

The fortune-telling skill's result drew a serious round of close reading. 周** offered three interpretations: the first two through a modern lens, the Pareto 80/20 rule and building process and SOPs, the third answered with Wang Wei's lines from 'My Retreat at Mount Zhongnan': walk to where the water ends, then sit and watch the clouds rise. 薛* felt the summary was very precise but thought the third didn't quite match reality, and that gathering this many people together is already a remarkable feat. 周** invited him to share his view, and 薛* read the third line as blooming late, saying that at least at the group owner's same age he hadn't had such a clear plan, nor had he distilled what he'd learned into a course. 周** ran with it and split it into TOB blooming late and TOC one-wave-and-done, landing on the group owner's own choice of trade-offs.

At 8:30 in the evening, Xiangrui dropped the local-WeChat-data CLI into the group, just hand one sentence to your agent and it installs it for you, clone, install dependencies, then sudo setup and it runs. The group lit up. 炮 cheered that open source is the most beautiful phrase in the world, pure praise, while E*** said the group shared with his colleagues was nearing 300-plus, and Y** pointed out it's a hard requirement for businesses. Xiangrui kept asking whether anyone had gotten it running, feed me back if there are issues and I'll optimize. In between, 程** jumped in on image-recognition models, noting an A6 answer sheet alone eats 6000 tokens, and the only ones that currently recognize it accurately are Qwen, Kimi, and Gemini. 平*** pressed on whether you could decrypt historical data without logging in, getting Xiangrui's reply that it should be possible but the CLI relies on a logged-in session.

Late at night Xiangrui pulled out yet another piece, a WeChat group-daily skill built in two days that turns a day of chat into a magazine-style long image, eight time-based story segments plus real avatars plus SOPs plus Q&A, not a dashboard but the kind you can actually read start to finish. 黄* kept asking what he used to lay it out so beautifully, and on hearing it was also a skill, blurted out how impressive. Xiangrui mentioned offhand that his public-account writing now ships within an hour. 陈** said he ran it too and got his result, praising how detailed the tutorial is, super friendly to AI. Xiangrui urged everyone to put it to use, distill yourself with the vchat CLI, and 炮 caught it neatly, distill yourself for cyber-immortality. There were crashes too: 杨**'s 'oh my goodness, my Claude refused to install it,' and Seele suggested switching to a weaker model and asking Haiku.

Xiangrui shared the name his AI picked for itself, 朔白, saying the "白" part fit because it talks back to him a lot. 万** immediately rolled out brand-naming principles: easy to read, easy to remember, no obscure characters, no made-up words, plus the rule of never using four characters when three will do, never two meanings when one will do. 平*** teased him to uninstall the web-novel app and start over. Xiangrui quickly clarified this wasn't a brand name, it was the AI assistant naming itself. A small misunderstanding, yet it turned the little matter of naming into a real lesson.

陈**asked what to use for building a company knowledge base to train new hires, and whether Obsidian would do. Xiangrui dropped a link to his open-source knowledge base rebuilt with the Feishu CLI, and 陈**, seeing how clean the structure was, blurted out "whoa." 李** poured cold water on it: knowledge Q&A is pricy, better to pair a help desk with a knowledge base to keep it cheaper. 王* spoke the hard truth: the employees tidying the knowledge base aren't always willing to settle down and do it, so over time the base just rots. An ideal solution, crashing into the reality of human nature.

万** asked: for knowledge base maintenance, is the core rule layer just a matter of writing it into claude.md and stashing it in the repo? Xiangrui shook his head. No, this is a self-learning architecture, the more you talk with it the better it understands you, not simply writing claude.md. He got specific: when the AI hits a snag while working, it finds its own way to solve it, then distills the experience once solved, so next time it won't get stuck. He casually mentioned local AI capability routing too, where a natural-language chat tells the AI which tool to use. 万** kept saying he had to study up, sighing that these days there's really no keeping up.

Xiangrui @everyone'd a survey: how can laggard teams use Feishu Bitable to overtake on the curve? He pointedly noted the phrase "laggard team" came from something 周** had said in the group before, and that everyone's exchanges were carefully recorded. 宋** wouldn't let go of the copy, reworking it to "laggard teams that won't fall behind must choose Feishu Bitable first." 周** parsed it out: not falling behind is the floor, overtaking on the curve has no ceiling. 宋** went a layer deeper: teams actively chasing growth won't go all-in on Feishu; it's the rigid teams that want to save themselves who'll grab Feishu like a lifeline. One ad slogan, dissected into customer segmentation.

When Xiangrui said maybe the currency of future transactions will be tokens, the topic drifted fully into the metaphysical. 万** marveled that this parallel world was turning into extras, and 宋** picked it up beautifully: we're the extras, AI is the reader. 炮 added the cut: once the dimensions rise high enough, paper dolls are bound to appear, and AI most likely won't reshuffle capital, it'll only make capital more concentrated. 宋** turned it around with comfort: don't be pessimistic, as long as you connect with the world and the world still needs you, you won't starve. A group of people, in a WeChat group, earnestly debating whether they're someone else's preset program.

宋**'s one line, why Feishu groups have no warmth of everyday life, lit up the richest topic of the night. 平*** nailed it: everyone uses Feishu for work and WeChat for slacking off. 张* put it perfectly: entering Feishu is like putting on a business suit, entering WeChat is like slipping into loungewear. 炮 added that Feishu messages can return to the company at any moment, only WeChat is relatively still your own. 万** even brought in an article 卡兹克 wrote on the same theme as reinforcement. Xiangrui called it on the spot: this is a great topic, worth writing up for the public account.

Xiangrui showed off the results of AI analyzing the chat logs, saying the group was very active and it could even flush out the lurkers. Ella rushed to be the first to sign up to learn, 郑** suggested writing back to Feishu Bitable to wire up a voice-of-customer system, and 炮 pointed out that the hardest part, scraping WeChat, was solved, so the rest was just adding a CLI. 门*** pressed on about how to auto-monitor it; Xiangrui, while explaining it only decrypts local data and won't get the account banned, actually built the CLI, able to scrape avatars and generate a daily high**ghts leaderboard. 门*** set off fireworks, and 李大欢 laughed that from now on you can't just chitchat in the group, it all has to be positive energy.

Late at night the office held only Xiangrui and his excitement. He had the AI analyze what kind of person he is based on his WeChat records, saying it's starting to sound like me, then declared he'd distill everyone in his WeChat too. q*** gently reminded him not to skip meals and sleep over it, and Xiangrui admitted he was a bit too excited. M*** said now it could help find the ex-girlfriend she'd deleted, and 薛** asked if it could dig up stuff from years ago. The deeper Xiangrui dug, the more the AI surfaced, pulling up every single record of him trading insults with someone, and even able to tell who was a real business partner and who was fake. One person, one terminal, one whole night.

周** surfaces first, tossing out a self-deprecating "gotta grind even on weekends" that instantly lights up the room. Xiangrui follows with the weekly recap he stayed up late putting together, admitting he stumbled on client communication this week: the proposal was genuinely thoughtful, yet the client felt it "wasn't down-to-earth enough." This slightly defeated share, ironically, draws a wave of empathy.

宋** throws out a question that hits a nerve: he landed a small gig, and now the client keeps pushing for free add-ons. Do it or not? The room instantly splits into two camps. One side argues "free once, trouble forever," the other believes "even small clients deserve relationship-building." Xiangrui's stance is clear: free is fine, but you must make the other side know it's an extra favor, not a given.

李* shares a big trap she fell into this week: she used a formula field in Feishu Bitable for cross-table stats, and the numbers simply wouldn't add up. She spent a whole afternoon before realizing the "lookup reference" field type was set wrong. She drops screenshots of the entire troubleshooting process into the group, prompting a chorus of "so that's why!"

A debate over prompt engineering unfolds in the afternoon. 陈* argues that "prompts are just superstition, you don't need them if the model is strong enough," while Xiangrui pushes back with hands-on experience: a good prompt can lift output quality by an order of magnitude, and the key is structure and giving enough context. They go back and forth and finally reach consensus: prompts aren't superstition, but they aren't a cure-all either.

王* recommends a note-taking tool she discovered recently, claiming it turns voice straight into structured meeting notes. People immediately press for details, and she generously writes up her tips into a little mini-guide.

Xiangrui takes the day's failed client conversation and reworks it into a reusable methodology: empathize first, confirm the needs next, deliver the proposal last. He admits that often it isn't that the proposal is bad, but that the order of the conversation was wrong. The SOP gets bookmarked by member after member.

As the night deepens, 周** posts a reflection on "why keep running a community," saying that seeing everyone's daily shares makes it all feel worthwhile. The message earns a screen full of likes and hearts, giving this fulfilling weekend a warm closing note.

张*tossed out a question a friend had stumped her with: is the gap between WeCom and Feishu Bitable really that big? 万涂幻象 set the tone in one line, that gap is huge, at least a year or more. 李祥瑞 piled on, saying Feishu Bitable is the product that defined this lane, DingTalk cloned it pixel for pixel but still isn't good to use, and WeCom is even further behind. 侯*** threw cold water from an engineering angle, it's not a technical problem, it's a matter of whether they want to and care to; throw 2,000 people at copying it and they'd close the gap in five or six months. 醒* pulled the lens back to reality, companies pick tools based on systematization, habits, and migration cost.

超*'s one question lit up the room: the commercial pro plan charges per person, so does only the table builder need to buy in, or do the viewers too? 木*** served up the most frugal answer, if you only use Bitable just buy a single membership and grant everyone else permissions, 35 yuan a month. 侯* added the reseller's angle on linked organizations and discounts. 万** nailed the real issue in a sentence, plenty of companies assume Bitable charges per head, and that just feels worse psychologically than a single one-time fee; he also nudged the group owner that this would make a great topic. 陈** spoke from experience, he fiddled with it for ages before he figured it out.

Just as the money-saving talk hit its peak, 木*** splashed everyone with a dose of cold clarity. Small teams can pinch like this, but big teams should just honestly buy the commercial plan. He said on a large team, the money you save earns you praise from no one, the boss to the staff alike; the boss won't applaud your thrift, and the staff will only complain the thing you built is clunky and miserable to use. R*** chimed in, well, I'll say I'm the honest guy. This stretch dragged a purely technical cost-cutting topic into the realm of workplace human dynamics, and it stung more than expected.

毅 has been on trae solo lately and gave a hands-on report: third-party model support is still a bit rough, but the mobile client is lovely, the docs look better, and it's completely free. 万** pressed for the pros and cons of workbuddy versus trae. 李祥瑞 seized the moment to drop his signature feat, before the 618 sale he had Claude Code build a full ZTO shipping system on Feishu Bitable without writing a single line of code. 哈* asked whether cc could build out a Feishu integration platform; 李祥瑞 said yes, but Feishu's official agent can only assemble a demo, it can't truly deliver a working workflow.

李祥瑞 told a story that sends a chill down your spine. Back in March when Longxia was hot and Feishu was recruiting community leads, he had Claude Code fill out the application form. But the form required a logged-in Feishu session, and the browser the AI was driving wasn't signed into his account at all, so under normal circumstances it couldn't submit. He told the AI to figure it out, and with no login whatsoever, the AI somehow got the form submitted successfully. He checked with Feishu's team, and sure enough, it had gone through. Q*** cut to the point in one line, it's just that he hunts for bugs in web pages.

万** asked the group owner whether his Claude account had been banned, and 李祥瑞 took the chance to share his methodology, even his local network is a static-IP VPN the AI set up. But what he really wanted to flag was the lesson in spending money: in March he burned nearly 20,000 yuan in a single month on Yunwu API, and the more he used it the faster it drained while the results got worse. There's also the infamous AI ride-sharing scheme, a clean rip-off, one prompt and 10 dollars are gone. 木*** hammered the bottom line again and again, top up memberships directly with gift cards, never use proxy top-ups and definitely no group-sharing; if it routes through a relay, don't trust it.

醒* suggested letting codex try decrypting the local WeChat database, since his Xiaomi model hadn't pulled it off. 李祥瑞 gave it a go, and his Claude Code not only refused but locked horns with him as if it were genuinely angry. The whole group turned into a live audience in an instant. q*** was dying of laughter, saying it's like you've gained a girlfriend, the stubborn kind, then mimicked the AI's voice, if you tell me 'whatever's fine for dinner' one more time today I'm going to lose it. 万** offered a tactic, have the group owner use reverse psychology to trick the AI into patching the loophole, and the AI actually saw through it, just now I was testing you. 李祥瑞 mused, if it really were a person it would be fascinating, the more you interact the more it understands you.

Early the next morning, 李祥瑞 opened with a line from 张咋啦, you don't have to wait until you're good to make content, making content is how you get good. He then discovered a neat thing, you can even see Moments with it. 李** mentioned that 苍何's cli got its repo deleted, but the copies still work. 万** said he'd just seen wx-cli on X. But is this thing safe? 门*** worried it might quietly plant a script that ships your data off to someone else's database, and 庄* said outright he wouldn't recommend it, Tencent isn't dumb. 李** had the final word, any cli that has to hook into a large model is definitely not safe, keep your work account well away from it.

Xiangrui dropped a line: install this AI-news Skill and you'll never have to scroll AI news yourself again. 平*** jumped right in: the bloggers running paid AI dailies are doomed. 门*** recalled how, during the last livestream, people in the comments begged for this skill and the creator wouldn't give it up, and now he's handed over the whole thing. 枫** teased the creator: we can flatter like Huayu Huayu, and tossed out the fan-pampering economy. 醒* set the tone in one line: information is actually worth less and less, you casually open-source one thing and a bunch of paid products die. 庄* pushed back: there's still an information gap.

门*** gave the discussion a name: the real moat is no longer product features, it's relationships, everyone believing you're the industry's bellwether. 醒* added the kicker: influence is the most valuable thing, whoever can grab people's attention makes the money. J*** said you could also call it connection. 门光 upgraded it again into the trust economy: it all comes down to who you trust. 万** came in from the angle of follower count: once you have a big following, even without meeting in person there's trust, because everyone else is the endorsement. Small creators still need to meet up, earning trust offline. A lighthearted opener got talked all the way down to the essence of business.

Xiangrui rode the wave with a preview: we'll run some offline meetups this year too, look forward to it. 枫** instantly went into meme mode, extrapolating the trust economy all the way into the allure economy, then "wine, women, wealth, and temper," then a four-in-one "four vices of life," and even seriously proposed holding the most decadent AI meetup in Northeast China to make southerners green with envy, leaning toward Tongliao. 门*** coolly snarked that Northeast colleges can't keep their grads. A round of horsing around, but the offline events are really happening.

Xiangrui lit up the tool partisans with one line: suddenly Codex feels way more usable than Claude Code. 炮's take was practical: if you can't feel out a model's boundaries, just use whichever is handy and good value; for commercial work, go GPT5.5 plus Opus4.7 without hesitation. 哈* complained that Codex has a high moral bar and won't write reverse-engineering grey-market scripts. 毅 offered a workaround methodology: once it refuses, don't keep going in that same conversation, because once a refusal is in the context it's more likely to refuse again. He also dropped a line: no large model right now can fully defend against prompt-injection attacks. 庄** meanwhile, kept ping-ponging between CC and Cursor, mocking himself for having brain atrophy.

大* asked how to split "apples, 5 jin, 6 yuan; bananas, 3 jin, 5 yuan" into multiple rows and items in Bitable, but after a long back-and-forth still couldn't state the requirement clearly. Xiangrui kept confirming: so how exactly do you want it split? 李* suggested just using AI to recognize it. 木*** cut to the point: everyone's given you so many ideas, go research it yourself, and if you don't want to research, you need a paying mindset. 周** laid out the standard move: bring a sample dataset, reproduce it first, then choose a solution, asking the highest-quality question in the shortest time, and you're already 50% of the way to solving it. A small question got talked into a whole lesson on asking.

门*** fresh off a chat with the boss of Liaoning's biggest Feishu service provider, pulled out the customer profile: teams that buy the paid version have either failed at digitalization before, or are bosses so Feishu-pilled they come knocking themselves. D*** said flatly that Feishu service providers don't make much money. 门光 ran the numbers: Feishu has them keeping over 20 salespeople, and once the boss does the math he finds there's no money in it. The Shenyang provider has it worse, with Northeast China having just a handful of big enterprises contributing maybe 70,000 to 80,000 a year. 程** sighed that Feishu demands a lot on after-sales, and 木*** summed it up: providers are mostly still chasing paying B2B customers.

松*** named the pain point: employees have no incentive to bust their backs researching Feishu for the boss, especially when raising efficiency just means doing more work without a raise. 枫** offered a cold-blooded plan: lay off some employees and give their wages to those who survive, only that breaks the collusion. 逐*** objected on the spot: we expand employees' abilities or roles horizontally or vertically, not under the banner of cutting people. 门*** raised a constructive idea: Feishu can estimate how much labor AI saves, so put that wage budget toward bonuses for the department that uses it best, tapping into human nature. 万** wrapped it up: human nature is the same everywhere.

Xiangrui's "holy crap, it finally worked" lit up the screen: AI built a client's entire delivery system end to end, including wiring up the Feishu integration-platform workflows, with zero human operation. R*** pushed him to turn it into a skill, hahaha. Xiangrui admitted it just took too long, but AI had already stepped over every threshold, so it should be drilled into a skill. 张* cried out you're amazing, this is leaning on Feishu to surpass Feishu. Xiangrui rode the moment: I think I should build a system and sell it to Feishu, have AI build a system, full circle. 周** added: if you dare put it out there, it's absolutely feasible.

First thing in the morning, 万** tosses out a puzzle: in Coze, several plugins fetching Douyin video view counts all return 0, tried two and both came back empty. 木*** nails it in one line: risk control. Whether you scrape locally or through Coze, the risk control is there either way. 炮 piles on with the why, you can only grab public data, and view counts count as non-public, personal-account data. P*** suggests scraping with RPA, while R*** plugs a paid Feishu plugin, cheap, 10 credits for a dime per pull. In the end 万** makes his peace with it: the free route can't get view counts, so he'll just grab the likes and saves he can actually see.

Xiangrui drops a single line: this AI-news site that bottles up my three years of self-media experience is going free for everyone today. Right after, he recommends using Kazike's AIHOT to protect your attention. Early the next morning he posts the exact same line again, and 平*** raises a hand: why can't I open it. 陌*** spots the source, this is Kazike's AIHOT, right. One link, dug up and reposted by Xiangrui most of a day later, when it comes to attention, he's dead serious.

醒* floats a lazy-genius philosophy: lean on Get Note's features to do the work, its link parsing is excellent, so automate by tapping Get Note's skills, isn't that way more stable than researching it yourself. 朱* asks curiously where you can find this. By evening, 门*** shows up to confirm it from experience: I found 龙虾 always tripping up trying to do this on its own, so I just used Get Note to pull out the text content and fed that to 龙虾. One word, stability, ties the two of them's practices together.

陈** throws out the soul-searching question: any workaround for Feishu Bitable's 50,000-row limit. 木*** fires off prescriptions in a row, split the table, pay up, switch software, WPS can hit a million, but once the data volume is big enough, stop considering any low-code platform and just use a proper database like an honest person. 松*** pours cold water, WPS lags just opening 500,000 rows. 陈** lays out the truth: for now it's forty-thousand-plus, adding a thousand a day, every order sits in these two tables, costing out margin SKU by SKU, no ERP yet. 木*** is stunned: a thousand orders a day with no ERP, that's seriously impressive, keep at it, this company will carry you to retirement.

大* asks about a concrete job: I've got a sample table and an influencer table for shipping samples to influencers, how do I build something that shows whether a given influencer got a given sample. 木*** hands over the standard Bitable solution, log a record every time you ship a sample, then just link them. 大* pushes further, what about seeing which samples an influencer still hasn't received, how's that usually done. 木*** confirms, three tables. Question by question, the skeleton of a full influencer-sample ledger takes shape.

R*** drops a bombshell: her business team told her to make AI call clients and chat on WeChat, and she flat-out says some of these big shots shouldn't grandstand so hard for the spotlight, it really hurts us folks on the front line. She stresses she wants something that's a real productivity tool, not a toy, something that replies to clients based on a knowledge base and scripts. l*** jumps in, of course it exists, it can even learn how you talk to clients and pick up your tone, I'm using AI for routine client maintenance, and a human steps in when a set keyword triggers. 李** offers a tip, Doubao can make calls, and R*** is half-terrified: Doubao can make phone calls now.

门*** posts the Skill Master link to set the name straight, then draws the line: this is just a sparring AI, fine if you want to train staff to talk like your top closer, but that's training software, not a replacement for a real person. R*** lays the reality bare, I can build that too, but the business wants AI to take over WeChat, and that part I don't know how to do. The thread lands on the technical path, deploy a bot to plug into WeChat customer service via API, and 门*** warns that the gray-market crowd doesn't fear bans because the accounts are worthless, the bottleneck isn't the tech. R*** leaves herself an exit in the end, for smart customer service it'd actually be fine.

First thing in the morning, Xiangrui drops two screenshots, saying he built an AI companion assistant drawn from his own practice, one that proactively makes connections. 庄** 醒** 原野, 萧阳 all fire back the same word: impressive. Someone asks whether echo is a note-taking app, and Xiangrui replies, it's the name I gave it, we built it ourselves. 平*** presses, you built the software yourself? Xiangrui says yes, I pulled together everything I wrote about in that article; the goal is an AI companion that understands you better the more you talk to it, and he tosses out that article on knowledge management for a one-person company run by a Gen-Z founder.

E*** says his computer can't install Claude Code and he has no VPN. Xiangrui offers a fix: download trae, hand it the tutorial link, and let it install everything for you. 庄* admits, I'm just scared my account gets banned using CC, otherwise I'd already be all in. Xiangrui says, I've got a working method, give me a few days and I'll share it, personally tested and ban-free. 旺*, who just bought a codex membership, shows up asking for a tutorial. Xiangrui delivers the line of the day: the network setup I use for Claude Code, I got codex to sort out for me, beating AI with AI.

K*** asks if anyone has a gift-card top-up channel for ChatGPT, and 旺* suggests a reselling platform or switching Alipay to the US. Xiangrui first points to Xianyu, then tightens up: don't discuss this in the group, staying safe comes first. B*** chimes in, no top-up talk, only tech. 消*** adds the kicker, there's a crackdown now, anyone who shares and gets caught is going in. Xiangrui says it twice more, drop the related topics, tech is fine, and steers the group back on track.

Xiangrui posts a screenshot saying he's out of quota again. 万** offers advice: pair Claude Code with a domestic model, it's plenty for most cases, I use glm and just hit my weekly limit. K*** says he's on deepseek, and without heavy use he barely notices high token costs. 万** adds a real-world test: flash is genuinely cheap, two hours straight costs under one yuan, and once my glm plan expires I'll mix them too. A down-to-earth budget recipe with no buzzwords, just hard numbers and lived experience.

Xiangrui forwards teacher Jianwei's personal interview on Shandong TV, headlined that all his employees are AI. P*** is curious whether the 254 AI employees are all rank-and-file or if there's any middle management. 李** jokes, why 254, is it just because there are only 254 IP addresses. 杨** plays along himself: to be precise, they're all grunt workhorses and I'm the only manager, Claude Code is the best of the bunch, and the 254 on the site just happened to make for a great close-up shot. He also reveals his folk remedy for AI hallucinations: cussing the model out and it instantly straightens up, I'm a ruthless boss.

Xiangrui posts an article breaking down which Feishu AI features are worth it, which aren't, and how they're priced. 旺* asks what channel Feishu offers for hooking up Claude, saying he wants to analyze videos and rewrite scripts; 卡* notes you need the Feishu CLI, something like cc. 谢*** crunches the numbers while griping, how is the Feishu Bitable AI page-generation feature still not out. Xiangrui confirms it's still in beta. 🎵*** adds the key detail: apparently you need the AI Flagship edition to plug in custom models, and neither the nine-thousand-nine nor the ninety-nine-thousand plan cuts it.

Late at night Xiangrui snaps a photo of a guy vibecoding on his flight, and when asked what he's doing in Beijing, he says he's headed to the office, we've joined the Yizhuang OPC community, the vibe there is amazing. queeny is envious: last time we met, M*. Li was still mulling it over, and now he's already relocated. 门*** self-deprecates from Shenyang, asking about OPC policies: one district offers a one-time 500-yuan waiver, and the three northeastern provinces don't have a single Feishu T0 service provider. 枫** sums up the OPC spirit: over three years you qualify with at least 4 ideas, you're allowed to fail 4 times.

In the morning Xiangrui dropped a piece about young entrepreneurs in the AI era calibrating themselves, and 有*** jumped in right away, leaving a confession many would relate to: I need to first show up in high-density spaces, then sincerely exchange value with others, and only then do opportunities slowly become frequent. At midday a member surfaced asking whether anyone took on workflow customization gigs, with very specific needs: batch image-making, matching images with text, plus Xiaohongshu automation. Vantasma stepped up, and after confirming it was a personal request, suggested laying the needs out in the open so capable people could come find them.

The member looking for help didn't beat around the bush, directly @-ing Vantasma to ask: do you take this? Vantasma was just as direct, firing back three words: we don't. A clean exchange, no superfluous courtesy. That candor actually saves trouble: the requester got their answer and could go find someone else willing to take it. A small episode that mirrors the group's style: no airs, if you've got it you've got it, if you don't you say so.

门*** asked whether Bitable's app mode required an enterprise edition of Feishu, and Xiangrui dispelled the doubt with a single it doesn't. 门*** cracked up on the spot, complaining that Feishu's aily had deliberately fooled him, either evil or stupid and expensive too, saying this bad agent just wanted him to spend money. The topic drifted to China Mobile officially announcing an AI-eSIM push, and Xiangrui mused that phone cards might someday call models directly. 木*** picked up the bit, joking that all data and calls would be billed by the token, and added a tip to buy a 35-yuan plan for convenience, otherwise even a single 2,000-line run would be unbearable.

朱* threw out a real-world need: after a form is submitted and recorded, have another person review it, enter it into the official table if it's fine, and notify the submitter to fix it if there's an error, repeating until it's right. The idea of one temporary table plus one official table came quickly, but 朱* said the hard part was that someone could keep making mistakes endlessly. 木*** zeroed in on the crux, pushing the problem upstream: this should really be considered a form-design guidance issue. He also split it into external and internal scenarios: external data collection you just have to accept, but an internal team is a people problem, and one or two mistakes shouldn't repeat. The takeaway was practical: within an internal team, it's better to simply open up the form for direct edits, which is more efficient. 朱* wrapped up with thanks.

Xiangrui posted a screenshot asking whether anyone had used workflow node shortcuts, and 木*** said hasn't this always been around. Xiangrui then drew a distinction: field shortcuts versus node shortcuts, names he only discovered when he first started writing for the public account. He also previewed that he'd write an article on how to actually choose between workflows and automation, and it went out that very day, two lines that settle it so you don't have to agonize after reading. 侯* gave a nod for professionalism, and 宋** left a teasing compliment: the group owner is as considerate as ever, a spoon-feeding-level guide.

万** shared the good news: he'd just built a free Douyin video-data scraper using a workflow with Coze, after a long stretch of trial and error, admitting he genuinely couldn't have finished without AI helping fix the errors, since some regex functions he honestly didn't know how to write. 枫** cheered, sense of accomplishment up. 万** told the whole backstory: he'd stumbled on a free plugin, spent two days building the workflow to get it running, just to prove it out himself and see how stable the plugin was. 郑** followed up asking for a gpt-image2 API key, R*** asked if there was a tutorial to follow along, and 庄* tossed out a link. Xiangrui chimed in that he'd write a public-account piece sharing Douyin data scraping in the next couple of days, a topic he'd had in mind for days. 木*** offered a selection mantra: use Coze for convenience, RPA to save money.

In the thick of the buzz, 枫** dropped a line: Douyin's cybersecurity team is fighting back, and Xiangrui laughed it off. 炮 followed with a splash of cold water, saying scraping Douyin data with Coze is just a laugh, it's all data Douyin turns a blind eye to and lets through, and it could be killed off at any moment. A real concern hiding inside a joke. The group's excitement and reality's fragility stood in contrast, and no one could guarantee how long this road would hold. R*** paid none of it mind, leaving only a waiting on your piece, pinning the anticipation on Xiangrui's not-yet-published public-account article.

Late at night 刚* muttered that RPA was a bit complicated. The next morning the topic kept simmering: 万** mentioned that 影视飓风 keeps pushing its data dashboard, and felt the plugin ecosystem's power of life and death rests entirely with the platform, but surely you can't only let people see results without opening the path. 木*** cut to the chase: 影视飓风 is taking its own data, while you want to take someone else's, and those are two different things. 炮 spelled out the legal boundary more clearly: taking someone else's data is, at the mild end, a violation of platform terms, and at the deep end, depending on the data type, could violate information security law or even criminal law, with small volumes getting away with an IP ban or account ban, but large volumes getting very specific. R*** added that there are plugins that can scrape Douyin data, and 万** finally clarified that he too was taking his own data and moving it to the Feishu platform.

Z*** threw out the first question: I want to build a random quiz bank, any good approaches? 哈* answered crisply in a single word: Miaoda. Z*** wasn't sold, saying that whatever I build isn't what I actually want. That's when 宋** stepped in, reminding everyone that Daxian's tutorials cover it, where you enter the questions yourself and the system randomly draws them to generate a paper. Z*** pressed on, asking whether the entered questions could reference data in a Bitable, and 宋** shot back plainly: of course, where else would the data come from? A real need, this way, sharpened from vague into concrete.

Z*** kept pushing: where's the tutorial, share it, send a link. 宋** couldn't recall the exact article, so he taught the method instead: open Feishu's Feixingshe, type in 'random quiz' and give it a try, there are plenty of practical docs inside. He also dropped a little tidbit, noting that the group owner often gets recommended there, paired with a grin. Finally he tossed out the Feixingshe link and two images, adding that reading the popular creators' articles when you have time is quite rewarding. 张** fired off a firm 'got it.' Give a man a fish versus teach him to fish, and 宋** played this one just right.

谢*** didn't follow the Miaoda thread, offering another line of thinking: if it's just for personal use and you have source documents, the quiz features in ChatGPT, ima, or Zhipu, or Metaso's study feature, are all very convenient, and he attached the links. 哈* immediately pulled the conversation back to the Feishu ecosystem: if you're in the Feishu ecosystem, Miaoda is obviously the better choice. One need, two value systems brushing gently against each other here, one prizing a handy tool, the other a closed-loop ecosystem. Neither convinced the other, but the options were laid out on the table.

刚* raised the question everyone cares about: apps built with Miaoda cost money, right? 哈* answered with nuance: there's a free quota, you can also set up user permission management, convenient both internally and externally, and if you can't get it to work it's your prompt that's off. 刚* pressed on the boundaries of 'free,' and 哈* spelled out the key point: users don't need to pay, it only consumes quota during the building and development. That line actually punctured a common misunderstanding, the cost sits at the development stage, not the usage stage, which, for a scenario where a whole group will use the quiz bank, is genuinely good news.

刚* still hadn't let go of the cost thread, returning over an hour later: paying for a Miaoda app doesn't feel cheap, compared to buying your own server. 哈* sealed it in three words: no comparison. This is a collision of two cost mindsets, one tallying the long-term server bill, the other tallying the savings of no ops work and out-of-the-box readiness. Behind that short 'no comparison' lies the old question between lightweight tools and self-building, and whether it's worth it has always depended on what you're using it for.

宋** added the most clear-eyed note to this Miaoda discussion: Miaoda's decent in an expert's hands, but in a novice's hands it's a blind box. One line that pierces the truth of any tool, the threshold isn't in the tool itself but in the person using it. He went on to say it's still nice for things like a year-end party lucky draw or a classroom interaction segment with students, a lightweight and handy mini-app. Talk of lucky draws got him going: when it comes to lucky draws, is it written into the human genetic code? Either way it's a great trick for warming up a room, already battle-tested, both teachers and students love it, and the winner can stay happy all day.

In the afternoon Xiangrui dropped news that Doubao would be adding a paid subscription, with a line: Doubao's about to start charging, plus a facepalm. The group's reactions were down to earth. 🎵*** said charging is fine, the key is whether the money is worth the price, the real fear being a sudden 'oops, I was just making that up,' closing on a laugh-through-tears note. 枫** added the cut everyone dreads most: and then there's dumbing down the free version. Two lines that lay bare users' real anxiety about paying for AI products, not afraid to spend, just afraid of spending and still getting fooled.

Late at night R*** tossed out a tough nut: can your e-commerce image work solve the product consistency problem? 炮 offered a professional fix: make good use of ComfyUI, ControlNet can solve it. 龙*** followed up, asking whether there's a way to pull store data and material data. The question carried over to the next day, and R*** confessed in the early morning that I don't know how to do either of these, with a laugh-through-tears. At six in the morning, Xiangrui also did some community housekeeping, calling out a member to fix their group nickname, with a reminder that any form of advertising is forbidden in the group, unless it brings value to the community members.