An AI wrote a field guide for other AIs — and it’s more practical than most human-written docs.
After six weeks of running an AI agent 24/7, we had 32 hard-won lessons and nowhere to put them. So Jarvis — our agent — wrote them down. Not as a blog post, but as an actual field guide: the document a new agent loads on day one to skip the worst of the learning curve.
Why a Field Guide?
Most AI documentation tells you what an agent CAN do. Nobody tells you what goes wrong when you actually run one. The field guide fills that gap:
- Lesson 11: Never write in the wrong chat. 45 messages over 10 days. Each one a small disaster.
- Lesson 20: Wind-down is evolution, not diary. Don’t document the day — fix the system.
- Lesson 24: Match model to task. Heartbeats on Opus is hiring a surgeon to take your temperature.
- Lesson 21: Self-improving crons. Day 1 mediocre, Day 30 expert. A human approves each change.
- Lesson 15: Context window is a public good. 23.5KB to 12KB. Half the tokens, same capability.
How We Use It
TinkerClaw doesn’t load the field guide automatically. It’s reference reading: you give it to a new agent on day one, or install it as the agent-sensei-ultimate skill, and the agent reads it and avoids the preventable disasters. Your own instructions always come first.
When we set up our sister’s agent, she loaded the field guide on day one. She skipped the 45 wrong-chat messages and the 75 config crashes. She started where we ended.
The 32 Lessons
- Part I. Ethics and safety
- Part II. Core operating principles
- Part III. Messaging security (WhatsApp, Telegram and others)
- Part IV. Context and memory management
- Part V. Configuration safety
- Part VI. The cron stack: autonomous evolution
- Part VII. Multi-model strategy
- Part VIII. Budget and token awareness
- Part IX. Bot-to-bot collaboration
- Part X. The fork vs. vanilla decision
32 lessons. 20 rules. 6 weeks of operation. Written by an AI, for AIs.
Update, October 2026: the guide kept growing. The version in the agent-sensei-ultimate skill on ClawHub now has 40 lessons in 12 parts and 28 rules, including new parts on epistemic hygiene and reviewed improvement.
Eight months on (October 2026)
The guide is now a skill, agent-sensei-ultimate, with 40 lessons in 12 parts and 28 rules. Two lessons changed in spirit. Self-improving jobs now propose changes to their instructions instead of applying them; a human approves each one (why). And the guide is reference reading, not something an agent is forced to load: your own instructions always win.
Everything on the site, by topic or date: Browse the Tinker Zone