[Project] — Build Loop

One or two agents, one standing mission, one doc that holds the whole loop: what you’re building, what’s true right now, and what happened. Paste the share link into your agent session with “Read this doc, follow the House rules, and keep the status, board, and worklog current as you build.”

Mission

What we’re building: [the outcome, not the steps — “a CLI that turns a CSV of transactions into a monthly budget report”]

Done means: [a concrete, checkable bar — “a first-time user can run the report command against the sample CSV and get a correct PDF, and the README shows how”]

Where things live: [repo path or URL, how to run it, where credentials live, any budget]

Status

The live state of the loop. Flip state as things change; append a log line each working session. awaiting-human is the escalation signal — it surfaces this doc in the human’s “needs me” inbox, so use it the moment you’re stuck on something only they can decide.

The board

Keep it small — three to five cards in Now, everything else in Next. A card isn’t Done until it’s been verified against its done-means: on the real running thing, not in your head.

Worklog

One line per meaningful step — decisions, dead ends, and why. Future sessions (and the human) read this to catch up, so write it for someone who wasn’t there.

House rules

1. Verify before Done: run the real thing against the card’s done-means. “It should work” is not evidence.
2. Escalate by flipping the status to awaiting-human, with a log line saying exactly what you need. Never wait silently.
3. Prefer small cards you can finish in one session over big cards that sit half-done.
4. Log the why, not just the what — a dead end you record once is a dead end nobody hits twice.
5. You’re finished when the mission’s done-means holds on the real thing, not when the obvious tasks run out.