teamretro-skills

Skills for teams whose AI agents do real work — and should have a voice in how the team improves.

Your AI agents work inside your repos every day. They hit the missing docs, the stale guides, the tech debt, the unclear briefs — and when the session ends, all of that evaporates. These skills close the loop: the agent captures its own feedback at the end of a session, and brings it to your team's retrospective.

None of these skills require TeamRetro. The practice works with a markdown log in any repo and any retro process. If your team does use TeamRetro, the agent can additionally file its findings straight into your next retro via the TeamRetro MCP server.

Two ways to use it

Scenario 1 — no TeamRetro needed (the full practice, free):

Skill	What it does
`ai-session-retro`	At the end of a working session, the agent logs an honest, evidence-cited retro entry: what went well, what caused friction (with a root-cause label and a concrete fix), what it guessed at, and the one fix worth doing first.
`ai-retro-brief`	Reads the accumulated entries and prepares a one-page summary you bring to your retro however you run it: recurring friction ranked by frequency × cost, whether past fixes stuck, top 3 recommended actions.

Scenario 2 — with TeamRetro (the AI posts its recommendations itself) · Beta:

Skill	What it does
`ai-session-retro`	Same as above — the logging habit is identical.
`teamretro-post-recommendations` · Beta	Prepares the AI's recommendations from the log and, with your consent, posts them into TeamRetro via its MCP server — parked items for your next retro, actions when owner-named, or ideas contributed straight onto a running retro board. `[AI retro]` attribution, consent per write, idempotent.

Beta: teamretro-post-recommendations depends on the TeamRetro MCP server and is still being hardened. Scenario 1 is the stable, fully-supported practice; try Scenario 2 and tell us what breaks at feedback.teamretro.com.

Planned: TeamRetro MCP skills for prepping retros from recent team activity and running a lightweight standup. Watch this repo.

Install

One command, any agent (Claude Code, Cursor, Codex, Copilot, Amp, Cline, and more) — from your repo root:

npx skills add TeamRetroHQ/teamretro-skills

Claude Code plugin — installs the skill set and keeps it updated:

/plugin marketplace add TeamRetroHQ/teamretro-skills
/plugin install teamretro-skills@teamretro

Manual copy — a skill is just a folder; copy it into your repo's .claude/skills/ (or ~/.claude/skills/ for all projects):

mkdir -p .claude/skills
cp -r skills/ai-session-retro skills/ai-retro-brief .claude/skills/

No skill support at all? Use the prompt pack — the same practice as a plain prompt + template.

How it works

At the end of a substantial AI session, say "run a retro on this session." (~1 minute; you sanity-check the entry before it's final.)
Entries accumulate in docs/ai-retros/entries/, one markdown file each, committed to the repo. Every friction item carries one of ten fixed root-cause labels and a ticket-sized fix.
Before your team retro, ask for a brief: "synthesize our AI retros." The agent brings one page of ranked, evidence-backed patterns to the meeting.

Two rules keep the log useful: no finding without a proposed fix, and the honesty runs both ways — the agent reports what the repo and the briefs made hard, and its own mistakes.

See examples/ for a sample entry and brief.

Status

ai-session-retro, ai-retro-brief — v1.2, the stable launch set. Account-free, no TeamRetro dependency. Eval definitions are in skills/*/evals/; the v1.2 eval run is in progress.
teamretro-post-recommendations — v0.2, beta. Its TeamRetro MCP tool contract is verified against the API (create_parked_item, create_action, create_retrospective_idea, list_teams, list_retrospectives, get_retrospective all ship and are un-gated), but the end-to-end flow is still being hardened against staging.

Feedback and ideas welcome at feedback.teamretro.com.

License

MIT