Zhang Pinpin · Hands-On AI Coach
← Back to home

One person, one agent: my AI community dashboard

Long-Form Video · EP0078 August 3, 2026 03:27
What this episode covers

Grown purely organically, we're at over 400 people, close to half of them active, and 308 people have reached out on their own. Daily messages went from two-hundred-something to nearly 500.

Of the people who sent a direct message, roughly 40% showed a clear buying signal—those are the ones I take over and follow up on myself.

Companion files · Drop them into Claude Code
WeChat Work AI Bot · Architecture Notes (the actual page from the video) HTML 36 KB
The actual page shown in the bottom half of the video, as a self-contained single file. The whole picture of one callback pipeline plus a row of scheduled jobs (message sources → entry layer with TLS, a resident process and full raw logging → orchestration with triage, knowledge-base retrieval, a live-search top-up and the answer itself → side jobs for the daily digest, knowledge increments, health probes and voice keep-alive). Plus six rules that have not changed since day one (standard library only, no framework · log first, features later · fail-open on features, fail-closed on anything sensitive · state is just files · draw the human/machine boundary · monitoring must not run through what it monitors). Plus the seven-gate chain for group Q&A, built on the premise that for most messages the correct behavior is to stay quiet. Plus the six modules, the stack, and seven hard-won lessons (don't pin logic on a platform's field formats / don't use JSON for long Chinese prose / knowledge staleness can't be solved at write time / "sensitive" also covers promises made on your behalf / always run a control group when changing a judgment prompt / verify from the user's side / never blindly retry a call that costs money or has side effects). Ends with a six-step opening sequence for handing the page to your own Claude Code. Credentials, server details and full configuration are deliberately left out.
Community Operations Dashboard (shareable version) HTML 61 KB
The shareable version of the dashboard from the opening of the video, with the data inlined so the single file opens on its own. 53 days of real numbers: 17,488 messages logged, 12,846 handled by the agent, 265 friend requests auto-approved. It covers daily message volume with a 7-day average, monthly rhythm, time-of-day distribution, a day-of-week × hour heatmap, community composition and participation rates, user tiering and personas with head concentration, growth and retention (first-time speakers, still active, gone quiet), agent performance (intent mix, image understanding, human takeover), the users who ask the most, plus high-intent leads and funnel gaps read out of real direct messages. Every member and client nickname is anonymized, and direct messages are only aggregated—never quoted.

The video opens with a disclaimer card I put there myself: this content is for AI learning and sharing only. The official tools are the platform’s own built-in assistants, or a service account once your company is verified. What follows is a framework and a line of thinking, not a tutorial.

Running a community group has probably never been this easy

This is the operations dashboard my AI built for me.

The group is basically in a state of organic growth—driving engagement, answering questions, following up on leads. Grown purely organically, it is over four hundred people, close to half of them active, and 308 people have reached out on their own.

The dashboard makes it obvious how active it is: daily messages went from two-hundred-something to nearly five hundred.

Of the people who sent a direct message, roughly 40% showed a clear buying signal—those are the ones I take over and follow up on myself.

What the dashboard actually shows

  • The daily message-volume curve, the active time-of-day distribution, and a day-of-week × hour heatmap—which, to some extent, actually guides when I publish my videos;
  • The community can be segmented and split: a fans group, long-term clients, a partners group—and the operating strategy and activity metrics are different for each;
  • It also does user tiering and personas;
  • Who the group’s active contributors are;
  • The distribution of question types the agent answers;
  • Who asks the most;
  • Where the latent business opportunities are;
  • And high-intent direct-message leads, plus the window in which they close.

Why people keep asking about the framework

Friends in the group all feel the agent replies very naturally and at high quality, so they all ask me how I designed this agent framework.

Back in the workflow era, roughly two years ago, I had already built a complete intent-recognition knowledge base for a medical setting, along with natural-language customer-service logic. That medical setup is genuinely complex. My community is an AI community—it does not need that level of intent analysis, long-term follow-up, or per-client knowledge-base separation. What runs here is a simplified version of it.

Why I’m not sharing the source code

The reason is simple:

In the age of AI development, rather than giving you source code, it’s better to give you a framework you can learn from and a line of thinking. That way your own community-operations system gets rebuilt around your needs, instead of being a clumsy imitation of mine. Plenty of small details in there you can redefine for yourself.

And if the jargon on that page is over your head? No problem—throw the web page at your own agent, Claude Code or Codex. It will keep working through it with you and build the whole thing out.

Both pages from the video are attached at the bottom of this post: the architecture notes (the one to hand to your own Claude Code—it ends with a six-step opening sequence) and the operations dashboard (shareable version, data inlined, opens as a single file). Both are self-contained, with credentials and server details deliberately left out.

How it’s put together

Authenticate first. Logically, the first step is user authentication—different users talking to the agent genuinely do get treated differently.

Then answer. When it replies, on top of using a good model:

  • it injects knowledge from my own past replies in the group and the content of my blog;
  • it runs one round of search through Perplexity’s online API to fill in the latest news;
  • it deliberately imitates the way I express things;
  • and depending on what the question needs, the answer may be short, fairly involved, or hand over a blog link for further reading.

Side-channel operations are bolted on separately:

  • 500 messages a day and some people have no time to read them → it produces a daily digest;
  • my own words in the group periodically go back in as knowledge-base increments;
  • it runs self health-checks and keep-alive;
  • and there’s one genuinely fun feature I call over-the-top compliments—using my own cloned voice to praise people in the group in flowery, situation-aware ways.

The rules matter more than the code

Through development I give the agent rules: reply logic, the human/machine boundary, and so on—including gate design and sensitive-word detection. All of it grew and iterated one step at a time.

The page also lists the pitfalls that are easy to fall into, which models and libraries I used, worked examples of the actual configuration, and some further reading.

One last thing

I don’t know whether today’s content will even be allowed through. Either way, I had my Claude Code prepare carefully for it—and if it can’t go out, it goes back to being group-members-only.

Running a community group has probably never been this easy.