Medical background → cultural-sector day job → self-taught AI workflow builder
Can't write a for-loop. 46 public repos anyway. All born from conversations in Chinese.

🇨🇳 中文版 · 🏠 aliceljy.github.io · 🎮 playground

_{29 active public repos · 17 retired with honors · 27 private · 132 merged PRs, 109 of them in other people's projects · 600+ articles · numbers as of 2026-10-22}

Jan 2026: "Hey Claude, help me write a WeChat article." (Claude Desktop)

Feb 2026: "OK maybe just one bot..." (still on Desktop)

Mar 2026: Typed claude in terminal. Claude Code era begins. No turning back.

Apr 2026: Codex shows up. Now there are two of them. They review each other's code while I watch. By month-end the roster peaked at five:

  ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌────────────┐
  │ AntiBot  │  │    睿智   │  │ AWS-bot  │  │  小试AI   │  │LanceDB Test│
  │ Opus 4.7 │  │Gemini 3.1│  │Codex 5.5 │  │MiniMax M2│  │  Opus 4.7  │
  └──────────┘  └──────────┘  └──────────┘  └──────────┘  └────────────┘
                       Peak roster — Apr 2026

May 2026: An "E-twin" — a digital me on Telegram. Eerie but useful. Also: the Hermes consolidation. 睿智 and 小试AI merged into Anti. Five bots collapse into three. Same throughput, half the moving parts. (小试AI stays on as a pen name — it'll outlive us all anyway.)

Jul 2026: The EC2 goes. Keeping a server alive just to run cron stopped making sense. Three become two.

Sep 2026: The OpenClaw chapter closes on the 14th. Anti, LanceDB Test and the local Docker bot stack shut down; my OpenClaw repos are archived as a record. A week later Claude Code and Codex leave my Telegram bridges, and Claude moves to the official Remote Control. Same month, the playground opens: a repo for things built purely for fun.

Oct 2026: Grok pulls up a chair at the review table. Codex nearly left in the middle of the month, then stayed: reviews, code and images go back on its desk. Meanwhile my prompt folder grows up into an actual system with a front desk and a bouncer (see the Arsenal).

Now:

  owner         whoever I'm talking to, usually Claude Code
  reviewers     Codex · Kimi · AGY · Grok   cold reads, kept apart from each other
  consultant    ChatGPT on the web      for "is this reasoning actually sound?"
  night shift   Telegram bridges (Kimi ×2, AGY ×1) · E-twin
  Discord bots  0                       (peak was 5. we don't talk about April.)

I still can't write a for-loop from memory. Claude doesn't seem to mind.

Class:         Non-coder with suspiciously many repos
Superpower:    死磕 — will bang head against wall until wall breaks
Weakness:      Social situations, verbal expression, for-loops
Fuel:          High-protein everything. (No fat, I have standards.)
Hobbies:       Singing, swimming, laughing (at everything, really)
Temperament:   Impatient optimist — breaks things fast, bounces back faster
Hidden gene:   Whatever makes you mass-produce repos at 3am and still function
Side quest:    CortexReach member · 109 PRs merged into repos that aren't mine
Partner:       Claude — Desktop first, then Code. It runs a crew now; I still talk to it most.
Terminal:      claude → ^C^C → claude (that's it. the full skill tree. maxed out.)

Stay optimistic, funny, curious, and a little bit lucky. Then dive in and don't stop.

The old map stacked 11 floors like a tower, and every repo got filed wherever it fit. The new map follows the work instead. A request comes in, the crew picks it up, draws on what it knows and on what it knows how to make, ships something, and whatever we learned flows back in. Every layer answers exactly one question, and every repo sits on exactly one layer — the one you'd check first when looking for it.

        me ── talking in Chinese, mostly from a phone
         │
         1  REACH ──────── where do I talk to it?
         │                 terminal · browser · Telegram · WeChat
         │
         2  CREW ───────── who does it, and who checks it?
         │                 one owner + four cold-reading reviewers
         │
  ┌────> ├── 3  MEMORY ─── what do we already know?
  │      │                 history · decisions · research
  │      │
  ├────> ├── 4  ARSENAL ── how do we make things?
  │      │                 prompts · thinking lenses · skills
  │      │
  │      5  FACTORY ────── what actually ships?
  │      │                 articles · videos · slides · apps · toys
  │      │
  └──────┘  lessons flow back into 3, good recipes into 4

  ══════════════════════════════════════════════════════════════════
         0  GROUND ─────── what keeps it all running?
                           harness · two Macs · disaster-recovery vault

_{6 layers · 1 feedback loop · 0 for-loops}

Both directions count. Most of these carry my words to the AI; one of them starts the conversation itself.

One rule runs the whole crew: whoever I'm talking to is the owner. The owner plans, hands work out, reads the results back, and brings me the disagreements instead of smoothing them over. I never carry context between AIs; at most I paste a path.

It forgot my name once. I took that personally. Now there are three tiers: distilled conclusions (RecallNest), raw transcripts (Deja), and researched knowledge (hippo). Want a decision → RecallNest. Want the exact conversation → Deja. Want what the literature says → hippo.

Memory is know-what. This layer is know-how: recipes, methods and skills the crew reaches for when it has to actually make something. Headliner first.

It started in April as "save this prompt, it's pretty." By October it had a front desk, a filing system, a bouncer, and a librarian who answers in three hops.

 SCOUT    a Grok-powered scout drops candidates every morning; I toss in links too
   │
 GATE     0  do I already have it?        ← the busiest checkpoint
   │      1  is it actually a prompt?     an article about prompts is not a prompt
   │      2  what does it deliver?        image · video · interactive page · method
   │      3  which shelf?                 by what you want to do, never by style
   │      4  are the credits complete?    author · source · repo · date · commit + license
   │
 SHELVE   one recipe, one copy; other shelves get a pointer, not a duplicate
   │
 FETCH    "find me an XX" → index → purpose sub-index → one-shot-prompt.md
   │
 DRAW     Codex by default, Gemini via AGY as backup, Grok when it's urgent
   │
 SHIP     article art · posters · short videos · 3D pages

House rules:

• File by purpose, not by style. Style is the prompt's own business; purpose is how I'll look for it at 11 p.m.
• One recipe, one copy. Two copies drift. They always drift.
• "All duplicates" is a perfectly good result. One batch: 117 volumes in, 111 bounced. The bouncer has standards and no quota.
• Tested or not, say so. 194 of the 238 folders state their test status in writing, and "not test-generated yet" is a perfectly respectable answer. Vibes are not evidence.
• No credits, no shelf. Every recipe carries five header lines: author, source post, official repo, date collected, upstream commit and license.

Why it's private: most recipes are adapted from other creators, some under non-commercial licences. Keeping their credit straight is my job; redistributing their work isn't my call. So the shelf stays private, and the method is right here.

Writing — 600+ articles on 我的AI小木屋 since January

Video — the fastest-growing wing

Slides

Apps — screenshots in, structure out

Toys — built because I wanted to

Private on purpose. Nobody visits a building for its foundations, but nothing stands without them.

我的AI小木屋 (WeChat) · A non-coder's real journey building with Claude
Weibo: 安闲静雅