The wave is coming either way. This is how you ride it instead of drowning — told by a non-coder who shipped ~30 real AI apps without knowing what npm meant, scars and all.
Each note follows the same honest arc: the learning → the new thing I tested (a hook, a model, an MCP server, a skill, a factory part) → the frustration that bit → what worked and what didn't. The hook is always welded to the proof.
HOOK · ENFORCE I wrote the rules down. It broke them anyway.A rule in a file is a wish — the thing that stops the mistake runs outside the AI. ✓ worked MODEL · REVIEW Three other AIs found the bug mine swore wasn't there.A model reviewing its own work brings its own blind spots to the inspection. ✓ worked AGENT · FACTORY The AI knew all ten steps. The bottleneck was me.Swarm the jobs that don't depend on each other; sequence the ones that do. ~ mixed MCP · CONNECT “Check my Notion.” It can't. It's in an empty room.A chat model has no hands until you plug tools in — MCP is the standard plug. ✓ worked SKILL “Pull the claims, dedupe, filter, format…” again.A procedure you re-type from memory is one you'll eventually get wrong. ✓ worked HOOK “I'll remember to check next time.” You won't.Good intentions don't run on a schedule. Hooks do. ✓ worked TOOL Copy out, paste back, repeat until it breaks.An AI that lives in your project beats a chat tab you shuttle files through. ✓ worked MEMORY “We use X. Never Y. Deploy like Z.” Again.Standing instructions belong on disk, not in your morning re-briefing. ✓ worked MEMORY · FACTORY “Which one is right?” Same fact, three files, three answers.Duplicate a fact for convenience and you've scheduled a future disagreement. ✓ worked TECH “I have a working script. I just need a URL.”Sometimes the right tool is the small one. Flask turns a script into an endpoint. ✓ workedAlmost none of this is writing code — it's making connections: model to tool, app to platform, your data to the thing that acts on it. So the evaluations are the product. Every platform, model, and connector we tried — winners, scars, and the ones dropped on sight.
VENDOR · MODEL What we tried. What kept working.Vercel & Supabase scars · DigitalOcean & GitHub wins · Microsoft the worst · why Hermes keeps getting rejected · which models actually connect (OpenAI great, Claude weak) · the little things nobody tells you. open the ledger →Longer build-alongs — the full architecture, every scar, and working code you can run tonight.
TELEGRAM · REMOTE Drive the factory from your phone.BotFather & how bots work · polling vs webhook · the bridge build · wiring it to your machine + the droplet · 3 real connections · every hardening scar · Codex & Tailscale remote control · a runnabletelegram-bridge.js.
read the guide →
REVIEW · MODELS
The bug your model swears isn't there.Why same-lineage self-review shares its own blind spots · a committee of robots that don't share a brain · the real scars (Gemini looping 40×, the flaky genius, presence≠structure) · a runnable cross-review.mjs for your next git diff.
read the guide →
PROMPT · PERSONA
I built an AI that refuses to help you.Meet Wanda — the govt call-center worker who hangs up on you. Why a system prompt is a character bible, not a rulebook · the 5 moves that make her real · the sentinel-string trick · steal her whole prompt + a "build your own character" recipe.
meet wanda →
LLM · UI CARDS
Describe a workout. Get a playable card.The pattern behind generative UIs: the model fills a schema, your trusted components render it — never writes UI code. 6 composable blocks · tool-use forces valid JSON · clone the working app (forge-oss, MIT).
read the guide →
The field guides tell the story; the goodies are the tools themselves — real things I built and use, extracted clean so they run out of the box (bring your own key, never a repo, never a secret).
FREE · TOOLS Build the tool. Keep the tool.The whole shelf — Wanda & Forge (live OSS) · the cross-review runner & Telegram bridge · a CPU dictation clone · an adversarial model-debate harness · drop-in widgets & utilities. Each its own card, honest ship-status, code or a build-it-by-talking recipe. open the shelf →The field notes are the story. The classes are the hands-on build — the same scar, taught as an interactive lesson with a teach-back gate and a real factory artifact you keep at the end. Gated; not published yet — you're previewing.
🔒T4·01Rules Don't Enforce Themselves — the flagship cold openQA ✓ 🔒T3·01Three Other AIs Found the Bug Mine Swore Wasn't ThereQA ✓ 🔒F30An Agent, a Subagent, a SwarmQA ✓ 🔒F31Claude Code — the AI inside your actual projectQA ✓ 🔒F32A Skill — a saved procedure you trigger by nameQA ✓ 🔒F33A Hook — the automation the harness runs for youQA ✓ 🔒F34An MCP Server — the standard plug for new toolsQA ✓ 🔒F35The Constitution the AI Reads on Arrival (CLAUDE.md)QA ✓ 🔒F40Keeping Memories — one fact, one homeQA ✓ 🔒10Flask — a tiny Python server when that's all you needQA fixThe by-profession course for Claude Cowork — for the non-coder who wants to delegate real computer work, not learn to program. Same interactive template, same brand, same receipts. Structurally the bigger audience.
status: README + roadmap only · 0 modules produced