learning-from-chats

Learning from chats

Trigger

A recurring correction or preference has shown up across sessions, or the user asks to "learn" how they work.

Workflow

1. Gather signal: review recent chats/corrections for patterns — things the user repeatedly asks for, fixes, or rejects.
2. Distill each into one durable statement of intent. Drop one-offs and project-specific noise.
3. Classify each via the decision test:
   • Persistent constraint, no steps → rule (or an AGENTS.md line if cross-tool baseline).
   • Invokable procedure → skill.
4. Author with rule-author / skill-author. Generalize: replace project specifics with placeholders.
5. Sanitize: never capture secrets, client names, or internal hosts into a public asset.
6. Register and run loadout doctor.

Guardrails

• Capture intent, not transcripts. One clear statement beats a pile of quotes.
• Confirm with the user before promoting a contested preference to an always-on rule.

Pairs with

• skills: skill-author, rule-author
• rules: no-secrets-in-code