Get started in ten minutes

Everything before this chapter was the case, the schema, and the facilitation. This chapter is the doing. Ten minutes from now you can have the skills installed, a trigger that fires without anyone remembering, and a clear picture of what happens after the first five entries land.

Install the two skills

The practice ships as two free skills in the teamretro-skills repo, open source under an MIT license:

• ai-session-retro writes the end-of-session entry: the four-part, evidence-cited retro this guide has described, with one fixed root-cause label per friction item and a ticket-sized fix on every finding.
• ai-retro-brief reads the accumulated entries and synthesizes the one-page pre-retro brief: recurring friction ranked by frequency and cost, the root-cause distribution, whether past fixes were actually adopted, and three recommended actions.

Clone the repo and drop the skills into your Claude Code setup. Not on Claude Code? The repo’s tool-agnostic prompt pack carries the same schema (same four-part entry, same fixed labels, same “no finding without a fix”) to Cursor, GitHub Copilot, or any agent you can hand a system prompt.

Make the trigger deterministic

Here is the one setup step teams skip and regret. Skill invocation is probabilistic: if nothing triggers the retro at session end, the log starves silently and your first brief never has enough to say. The fix is to layer a deterministic trigger on top. Pick at least one of these; two is better.

A session-start hook. Add a SessionStart hook to .claude/settings.json that injects one line of context: “at the natural end of a substantial session, run the ai-session-retro skill and commit the entry.” Blunt — it fires on two-minute sessions too — but it’s the only trigger that depends on nobody remembering anything.

A /retro command. Create .claude/commands/retro.md with a short instruction to run the skill on the current session, exactly as specified. Typing /retro at the end of a feature branch or a triage shift invokes it on demand. Deterministic once typed; the weak link is human memory, so pair it with one of the others.

A convention line. One sentence in the scope’s context file (CLAUDE.md / AGENTS.md), or in the team SOP for non-code scopes: “End every substantial AI-assisted session by running ai-session-retro and committing the entry.” Zero infrastructure; treat it as the norm layer, not the trigger.

Exact copy-paste snippets for all three live in the repo’s SETUP.md.

What your first entry will look like

Run one session with the skill on. At the end, the agent writes something like this: a real shape from a marketing scope, not a coding one, to make the point that the practice isn’t dev-only:

# 2026-07-16 — Monthly ads account review

**Session size:** ~25 turns, 1 deliverable
**Outcome:** shipped

## Friction
- **[missing-context]** The brief didn't say the French campaigns were
  deliberately paused for Q3; ~40 min (est.) auditing a "broken"
  campaign that was fine.
  → Fix (altitude: process): add campaign-status flags to the monthly
    brief template

## Do this first
Campaign-status flags in the brief template — kills the largest single
time sink this month.

One label per finding, evidence with a cost, a fix sized to become a ticket. That first read is usually the moment the practice clicks: the agent names something about your work you half-knew and never wrote down.

After five entries, run the brief

One entry is an anecdote; five is where patterns start to mean something. After a week or two of normal work, run ai-retro-brief. It reads every entry, ranks the recurring friction, states plainly which root-cause group dominates (briefing, documentation, the work material, tooling, or the agent itself), and checks whether the fixes past entries proposed were actually adopted. The output is one page, written for humans walking into a retro.

Then close the loop, and there are two ways to do it:

• With any retro tool. The brief is just a document. Read it before the meeting, add the top item or two to whatever board you run, sticky notes included. The whole practice works this way, end to end, without TeamRetro or any other specific product.
• With TeamRetro, the agent posts for itself. If your team runs its retros in TeamRetro, the agent can prepare its recommendations from the log and, with your confirmation on each item, file them through the TeamRetro MCP server: parked items queued for your next retro, or actions when a fix already has an owner. Every posted item is prefixed [AI retro], so the room can see which participant raised it: the agent’s findings arrive on the board as its own contribution, instead of you relaying them second-hand.

Where to go next

That’s the whole loop: install, trigger, five entries, brief, retro. If you want to revisit the reasoning behind any step (the labels, the honesty rules, the facilitation), the guide hub links every chapter. And if you’re sending this to a teammate who won’t read six chapters, send them how to gather feedback from your AI agents — the 10-minute version of this whole guide.