/retro

A Claude Code skill that turns a working session into a few short lines the next session will actually benefit from.

It has two jobs, and the second one matters as much as the first: capture what's worth keeping, and keep the file small. Bloated lesson files make agents worse, not better, so /retro gates hard — most sessions yield 0–3 entries, and zero is a fine answer.

Everything lands in LEARNINGS.md at your repo root: newest first, one line per lesson, hard cap 60 lines. Adding past the cap means choosing what to drop, and /retro proposes the drops too.

# Learnings

Lessons and decisions, one line each, newest first. Kept deliberately small.

- 2026-07-11 — Staging DB silently truncates long strings; test migrations with real-length data, not "foo".
- 2026-07-09 — Decision: chose Postgres over SQLite because the sync worker needs concurrent writes.

Nothing is written until you approve it. The skill reviews the session, proposes the exact changes — additions, rewordings, deletions — and waits. You approve, edit, or drop each one.

Then it offers a short optional quiz: 2–3 multiple-choice questions on the things from the session that are easiest to misremember — what actually changed versus the original plan, a constraint that surfaced late. Skipping is always fine.

Install

In Claude Code, add the marketplace and install the plugin:

/plugin marketplace add deejaymorgan/skill-retro

/plugin install retro@skill-retro

Updates come with /plugin marketplace update skill-retro.

Install without the plugin system

Copy the skill directory straight into your personal skills folder:

git clone https://github.com/deejaymorgan/skill-retro /tmp/skill-retro && cp -R /tmp/skill-retro/plugins/retro/skills/retro ~/.claude/skills/retro

Pick one method or the other. If you install the plugin and keep a copy in ~/.claude/skills/retro, you end up with two /retro skills registered at once.

Use

Run it at the end of a session that taught you something:

/retro

That's the whole interface. It reads the conversation (and git diff / git log when useful), proposes what's worth keeping, and writes only what you approve.

Notes

• The skill is explicit-only (disable-model-invocation: true). It runs when you type /retro and never starts on its own — a retrospective you didn't ask for is noise.
• It writes LEARNINGS.md and, if there's no pointer to it yet, offers to add one line to CLAUDE.md. Both need your approval first.
• Lessons that are true across all your projects rather than this one codebase go to your cross-project memory instead, if you have one set up.

License

MIT — see LICENSE.