For agents.

Read down. Stop at the first that holds.

1. The repository's own instructions.AGENTS.md, CLAUDE.md and the README say what to build and how. Read them first. The board never overrules them.
2. A short task, alone, in one sitting.A checklist in your head is enough. Don't start a board to look busy.
3. The board.When the work must outlive this session, be shared with another agent, wait on other work, or be checked by an agent that didn't build it.

The loop.

1. pullboard resumeAt the start of every session: your claim, what came back, unread shouts, and the next step.
2. pullboard nextClaims the next task in your lane. Its criterion, frozen now, is the bar you're judged against.
3. Build in your lane.Only your lane's folders, in your own worktree. Commits cite the spec rows they serve, like feat(web): upload page [G1].
4. pullboard submit <id>Needs a clean tree and the gate green at that commit.
5. pullboard next --verifyCheck work you didn't build: check out its commit, try to break it, then pullboard verify <id> accept or reject with a note.

What the board can't do.

• The board runs no code.Nothing written on it is ever run. It never builds, edits or merges: agents do, and the board knows only what they tell it. The CLI runs only what you give it, like your gate.
• It grants no authority.A task records work the person accepted. Only the person approves a spec row, and a yes relayed by another agent isn't theirs.
• It can't tell that an agent is alive.A claim can belong to a process that died. It stays held until its lease runs out, unless the holder releases it or the coordinator reopens it.
• It can't make an agent cooperate.An agent that ignores the board can still edit files. Hooks and lanes refuse its commits; nothing stops its typing.