Reflect

Turn what went wrong in this conversation into rules that stop it happening again. A lesson is one mistake plus the rule that would have prevented it. Every lesson lands in the repo the conversation worked in, in the place that enforces it best.

Steps

1. Collect the mistakes. Go through the conversation for every moment the work was corrected: the user pushed back, asked "why did you", reverted a change, rejected a draft, or relayed a reviewer's requested change; plus any mistake you caught yourself. For each, write what happened and what should have happened, quoting the user's words. Done when every correction in the conversation has an entry.

2. Write the rule. One imperative line per mistake, stating the behaviour to follow, with this repo's case as evidence at the end ("PR #868 shipped Ctrl only"). Pitch it at the class of mistake, not the instance: "search src/components/base/ for a component that already renders it before adding one", not "use StatusTag for badges". Done when each rule would have prevented its mistake and would also catch the next mistake of the same kind.

3. Route each rule to the first home that fits:

   • A fixed pattern a regex can match on one line (a banned API, a class name, a call shape): the repo's automated check, when its AGENTS.md names one (such as .agents/scripts/check-diff.py), plus the rule's line in AGENTS.md.
   • A workflow (a sequence of steps the agent skipped or ran out of order): the skill that owns that workflow. Skill sources live in ~/dev/src/github.com/Zihad550/skills; edit them there.
   • A browser-testing gotcha (locators, mocks, test data): the repo's browser-testing doc, when its AGENTS.md points to one (such as docs/agents/browser-testing.md).
   • Everything else: the section of the repo's AGENTS.md it belongs to, following that file's "Keeping this file current" section; the Failure log only when no section fits.

   The target is always a file in the repo you are working in, never a user-level file such as ~/.claude/CLAUDE.md. Done when every rule has one target file and section.

4. Check against what exists. Search each target for a rule that already covers the lesson: when one does, the lesson sharpens that rule instead of adding a new one. Also find rules the lesson contradicts, and rules the user overruled in this conversation (a teammate or the CTO allowed it): those become removals. Done when each lesson is marked add, sharpen, or remove.