all field notes

Field notes

Notes on agentic development — retry budgets, context windows, evals, and the messy middle.

2026-08-14Notes to Self: The Interview Between an Issue and a SpecI write issues as three-sentence notes to myself. That worked while I was the one writing the code, because I filled the gaps from memory without noticing. An agent can read every file in the repo and none of my recall — so the missing step is an interview, and its output is a spec.10 min
2026-08-07Worktrees: The Missing Primitive for Multi-Agent CodingGit worktrees are usually sold as a convenience for juggling two branches. Run several coding agents against the same repos and they become something else — a concurrency primitive, where the branch is the lock and the only hard part is cleaning up.9 min
2026-08-06A Public Repo That Still Deploys Real InfrastructureThe third and final piece on taking a home-lab project open source: the moment scrubbing a public repo turns into "how does this still deploy to my own NAS," and the private pipeline built to answer it.7 min
2026-08-06The Advisor Tool: The Question Code Review Can't AskTests and code review ask whether a change is right. Almost nothing asks whether it's complete — whether the request, as written, described the whole job. That third question is what a stronger model reviewing the transcript is unusually good at, and the terminal shows you almost none of it.7 min
2026-08-04The Versioning & Deploy Pipeline I Give Every RepoA concrete recipe for a tag-based versioning pipeline and the staging/production split on top of it — copy the snippets, or follow it end to end.16 min
2026-08-03The Git History You Can't Actually DeleteThe second piece on taking a home-lab project open source: what "rewrite git history" actually turns out to mean once GitHub is involved. Hidden PR refs, an unredactable Copilot comment, a false alarm, and a purge that came back five times bigger than scoped.10 min
2026-07-29The Boring 90%The first of three pieces on taking a home-lab project open source: rotating a leaked credential, discovering a docker-compose.yml was secretly production, and using AI to finally make a years-deferred rename cheap enough to just do.10 min
2026-07-12Trust, Doubt, and a Production Rate-Limit IncidentA production incident post-mortem where the debugging partner was an AI agent — the real story is when I believed it, when I didn't, and why that mattered.11 min
2026-07-03Running Claude Code as Persistent Infrastructure: A Containerised SetupWhy a laptop-bound Claude Code session stopped being enough, and the containerised, SSH-plus-tmux setup that replaced it — persistent config, scoped MCP servers, and one image for both interactive and CI use.17 min
2026-07-01Quieting 4xx noise in FastAPI + Logfire without losing your 500sA one-line APIRoute fix for a non-obvious interaction between FastAPI's exception handling and Logfire's tracing instrumentation.3 min
2026-06-30Your CLAUDE.md Is an Architecture Document — Write It Like OneThe patterns worth stealing from a FastAPI project whose agent-instructions file quietly became its architecture doc — layering, config quarantine, the DI trap that breaks transactions, and more.12 min