llm-session-retrospective

Ask for a retrospective of recent Codex or Claude Code work. The Skill searches the current repo, sibling repos, and local agent directories, then produces a report with a ranked list of candidate skills and optional starter stubs.

The Skill automatically scans docs, plans, design notes, reviews, and TODO files in the repo plus ~/.claude and ~/.codex. It classifies repeated decisions and failures into reusable patterns rather than writing a chronological summary.

Yes, by default it writes a dated markdown report to docs/retrospectives/ in the current repo, creating the directory if needed. It only writes into a shared agent-skills repo when you explicitly request cross-project harvesting.

Report mode produces only the retrospective, propose mode adds ranked skill candidates, and scaffold mode additionally generates starter skill stubs. The mode is inferred from your request and defaults to propose when unclear.

Keep a pattern private or local when it references internal services, exposes sensitive product or customer context, or is tightly coupled to one team or monorepo. Public skills should solve repeatable problems across multiple repos or sessions.