After-Action Review Tool for Engineering Teams
The same kind of incident keeps happening, and the review after each one stops at who missed a step. AAR Maker runs structured after-action reviews: the incident and post-launch review I used to facilitate in person, now software a team runs on its own cadence. Each review becomes an artefact the team keeps learning from, with the discipline a good facilitator brings to the room. I built and shipped it into a 25+ team organisation; of my three tools there, it’s the one that needed an LLM. The teams there still run most of their reviews through it, months after I left.
“[AAR Maker] turns a meeting Chris used to facilitate into a structured artifact the team produces continuously. Conservatively saves hundreds of engineering hours a year.”
Christophe Louvion, CPTO
How a review runs
A review moves the way a good facilitator would take you: build the timeline, find the contributing factors, and map each one to where it actually lives, whether that’s strategy, structure, process, rewards, or people. When the answer is “someone missed a step,” it asks what made that step easy to miss. You come out with a short list of actions, each rated by relative urgency, so the team fixes the constraint instead of trying to fix everything at once.
Blameless by default
Most of what goes wrong is systemic, not personal (Deming), so “who missed it” becomes “what made this outcome likely.” When someone made a call that looks wrong in hindsight, the tool asks what made it seem right at the time.Dekker’s local rationality.
1 Dekker’s local rationality.
What the team keeps
Run enough reviews and the tool reads across the whole set for the patterns a single review can’t see: the gaps in how the organisation is set up that keep generating the same incident. With your tools connected, it pulls the bug and timeline evidence in and writes the action items back out as tickets, so the follow-through sits in your tracker.
Try the open-source version
The version above runs inside one organisation. Separately, an open-source version, built on my grounded-forge, runs as a connector anyone can try at aar.groundedforge.com, and it’s listed in Anthropic’s Claude connector directory. Add it to Claude, Claude Code, ChatGPT or Cursor and sign in with GitHub. It runs the review from a library of 22 openly licensed sources, among them the US Army’s after-action review guide, the NHS Just Culture guide and the Columbia accident report. In a synthetic test (one invented incident, two runs each, with Claude playing the manager and judging), Claude with AAR Maker drew out 24 and 25 of 32 facts the manager held back; the same model with the same sources and no AAR Maker drew out 7 and 2. The write-up gives the setup and the limits.