Auto-extract recurring corrections from Claude Code history into CLAUDE.md (claude-reflect / `/insights`)

Parses your local Claude Code conversation history, identifies passages where you repeatedly course-corrected the agent (e.g. "stop using npm run X, use make X"), and proposes diff-ready additions to your claude.md file. The community plugin is claude-reflect (open-source, search GitHub by that name); the native form is the /insights slash command shipped in a recent Claude Code update — type /insights and it surfaces recurring patterns from recent sessions.

Compounding effect over weeks: each rule that lands in claude.md removes one recurring correction from your future sessions. One operator's experience after running /insights: "Claude itself told me I was doing too much extra work — here, take this rule." Cheaper than manual curation, and catches patterns you'd never notice yourself.

1. For the native path: ensure Claude Code is on a recent version, then run /insights from any project session. Review the proposed rules and decide which land in project-level vs. user-level claude.md.
2. For the plugin path: install the claude-reflect plugin (open-source on GitHub) against your CC install. It scans logs and surfaces correction patterns.
3. Skim the proposals. Promote the ones that match your actual policies; discard noise.
4. Commit the resulting claude.md diff so the team picks it up.
5. Re-run weekly or after any large new project. The set of corrections drifts as your stack changes.

The proposals are only as useful as your sample. If you've been short with the agent ("no", "stop", "fix"), the plugin will infer brittle rules. Run it after a week of well-articulated corrections, and pair it with positive-framing — write rules as "do X, that works" rather than "don't do Y" (negative-framing skills perform worse in practice).