← ClaudeAtlas

buildlisted

Start or steer a supervised task-loop run. Use when the user types /build, or asks to work through a plan autonomously and have the loop keep going until it is done or a cap is hit.
HeisenbergI8/claude-harness · ★ 0 · AI & Automation · score 72
Install: claude install-skill HeisenbergI8/claude-harness
# Build — the supervised task loop A loop where **code decides whether to continue and what is next; the model does the work.** `task-driver.mjs` runs on `Stop` and blocks the turn with the next phase as its reason. That block is the iteration primitive: the turn does not end, and the model is re-invoked with instructions it did not write. ## What the hook already did `build-trigger.mjs` fires on your `/build` message and **creates the run record itself**. You do not create it, and you do not need to remember to. That is deliberate. A skill that merely *tells* the model to create a run produces no run when it forgets — and a driver with no run releases forever, which is **indistinguishable from working correctly**. The mechanism sits in the hook so it cannot be skipped. ## Commands ```bash /build <label> start a run /build <label> --plan <path> start it already bound to a plan /build --status where is it /build --resume continue after a pause /build --halt stop it ``` **Any other user message pauses a live run.** The loop advances only while you are silent. This is a feature: a loop that reinterprets its objective from conversation is how a run quietly becomes something nobody asked for. ## Your job when a run starts ### 1. Size the task honestly, out loud **The loop can only drive planned work.** The cursor advances through `#### Step N.M` headings, so work with no plan has no phases and nothing