Run the build loop

Describe the change you want, and your agent takes it through four stages, one section each below.

Route the request

Describe the work in your own words, however vague:

I need to make the login test deterministic. Route this request.

The agent runs work-intake, picks the skill that owns the request, and tells you why.

Write the contract

Ask for a spec, a short document that says what done means, with a plan for how to get there:

Write a spec for making the login test deterministic.

The agent writes docs/specs/<slug>/spec.md and plan.md, each acceptance criterion paired with a check that can fail. Read the spec and approve it before building. For work that spans several specs, ask for a brief first with author-brief.

Run the loop

Point the agent at the approved spec:

Run the work loop on this spec.

The agent plans, builds, runs your checks, and sends the change to independent reviewers, pausing wherever it needs your decision. If a check fails, it goes back and fixes the build before review.

Before completing a change, the work-loop checks all active Parent intent:, Brief:, and Discovery: fields in each artifact the change adds or modifies. A typed pointer must name an existing target of the field’s type. A Discovery: that points at an intent must be written <kind>:<slug> and name a live intent of that kind; a research or notes path there is not checked. A body-only edit still selects that artifact for checking.

Close it out

When the change is merged, ask for the record:

Close this work out.

The agent runs close-work, writes docs/lifecycle/<slug>.md with what shipped, what did not and why, and updates the workspace. To keep a lesson for later sessions, ask for project-knowledge, which saves what this work taught the project into the repository.

Next, release the change.