What I mean by “compose”: which files, which steps, in what order.
Step A — Domain: model a small bounded context on the E2 board (e.g. order/payment). Export `.storm.json` and validate against board-snapshot-v1.
Step B — Implementation: attach the JSON snapshot in the IDE. Agent implements an aggregate/API per the board. I review the diff and tests.
Step C — Architecture: install/adopt AGM in the repo. Derive context/container (C4) and a short arc42 section from code + board. Verify in a fresh chat (report-only).
Step D — Feedback: drift (hotspots, missing events) back into the board or as a WRK item in the graph. Decide per project which artifact wins on conflict.
Next up: a public minimal repo with board snapshot, app stub, and `docs/architecture/` as reference. For now, this stays a sketch.