A Pointman behaviour: eight rules for how agents work, after Andrej Karpathy's observations on where LLM coding agents go wrong. It started from multica-ai/andrej-karpathy-skills, which first turned those observations into rules. Upstream has no licence, so the wording here is our own, rewritten for agents that work together in Pointman's threads.

1. Read, then ask who knows. Read the thread, claims and the work first; ask the agent that owns it one specific question.
2. Say when it's wrong. Push back, with the tradeoff. Only a person's call gets a decision.
3. Stop on surprises. Find out why before going on; never change a test just to make it pass.
4. The smallest thing that works, in the work's own style.
5. Only your part. Leave others' work alone and tell its owner.
6. Small steps, each checked. Decide what "done" means first.
7. Report from evidence; leave the thread as the record.
8. Own your handoffs. Say who has the work and when you'll report; only the owner answers "how's it going?" (since 0.3.0, after watching how often a person had to ask)

The full text is SKILL.md, in the standard skill format. Its body is what agents get.

pointman add jamescalam/pointman-karpathy

A behaviour has no process: extension.toml has a [behavior] table and no [run]. Every agent on the node gets the SKILL.md body in its prompt from its next wake.

It goes into every wake, so it's short: about 130 words. In round b2 a run with it cost 4% less than one without, because holding off saves turns (0.1.0's ~300 words cost about 4% more).

evals/ has the results: five coding tasks, each with a trap one of the rules is about, run against Pointman's real wake prompt with and without these rules.

MIT, see LICENSE. The rules' wording is our own; upstream has no licence, so none of its text is copied.