A fellow in the Commons
Chris Gagné
About Chris Gagné
Agile coach turned builder, living in Nelson. I help leaders find where their work waits, lately where AI has made the work faster but results haven't moved. I build my own tools for it and mentor leaders along the way. Open to work, and happy to talk through whatever you're building.
In the Commons
A free coach for blameless after-action reviews
Software teams need good after-action reviews after incidents. Even with the process I built from my time leading incident response product at PagerDuty, my teams struggled to run one well unless I was there to facilitate. So I built an after-action review coach that runs inside Claude or ChatGPT. My last client still uses it; their CTO says it “conservatively saves hundreds of engineering hours a year.” This is a free, open version. It’s blameless by default: it asks what made the outcome likely and what made a call seem right at the time. It draws only on openly licensed sources (the US Army’s AAR guide, the Forest Service, NHS Just Culture, the Columbia accident report), so it works for any debrief: a launch, an event, a programme. Try it on a real review and tell me what broke or was missing. Sign-in is a free GitHub account, and it logs only your username and which file it read, not your input. Happy to help with setup by private note.
Build an AI assistant that answers only from sources you trust
One of the challenges with AIs is that they seem to base their answers on the mode of their training data, not genuine thought leadership. Web searches give LLMs access to more language, not truth, and they often have a hard time telling the difference. I’ve built a “knowledge work factory” of sorts—akin to the software factories many companies are using now to write software—for doing complex research and synthesis. You pick the sources: books, guides, your own body of work. My tool then ingests, summarises, and indexes those in a way that preserves provenance all the way to the final output, and audits every claim back to its source. AAR Maker is built on it, and that’s still in use at my last client. My corpus is about 24M words or ~35M tokens, about 35x the maximum that can fit in a 1M context window. With its provenance-protecting indexing and summary system, I can run heavy synthesis queries that consider my entire corpus in about 500K tokens of context, in a single session.