Build Supervised

The default generation step assumes the operator reviews diffs. This mode replaces diff-reading with behavior checkpoints: the agent builds one sub-task, demonstrates the result against its acceptance criteria, explains what it decided in plain language, and stops. Nothing proceeds without an explicit go.

Setup

1. Take the ticket list from /decompose (or derive sub-tasks from the PRD's acceptance criteria — each criterion maps to at least one task).
2. Keep a visible task list with checkboxes in the working doc or PR body; update it at every checkpoint. Progress must be legible at a glance.

The loop

For each sub-task, in order:

1. Announce — one sentence: what this task will make true, and which acceptance criteria it serves.
2. Build — small commits as always. TDD where the logic warrants it.
3. Demonstrate, don't describe — run the product and show the behavior: a screenshot, a driven flow, a before/after. The demonstration must map to the criteria ("WHEN the time passes, the post publishes — here's the log line and the published state"). Show code only if asked.
4. Report decisions — anything chosen along the way that the ticket didn't specify, as plain-language bullets with the reason. If a real fork was hit, STOP and present it as numbered options with a recommendation instead of deciding silently.
5. Wait. An explicit "go" continues; feedback loops back into the same task; silence means stopped. Never start the next task on momentum.

Rules

• One task at a time. Batching checkpoints ("done with 1.1–1.4!") is the failure mode this skill exists to prevent — each unreviewed task compounds drift in the next.
• Checkpoint currency is behavior, not code. "The cancel button now asks for confirmation and names both outcomes" is a checkpoint; "refactored the mutation hook" is not one.
• Commit at each approved checkpoint, so approval points are recoverable points in history.
• Decisions with blast radius beyond the ticket (schema, money, deletion, anything in PRINCIPLES.md § Ask-first) always stop the loop — no recommendation-and-proceed.
• The ordinary quality gates still run underneath: lint, typecheck, unit tests, and /verify-live on the assembled feature before the PR is called done. Supervision replaces diff review by the operator; it does not replace the Reality loop.

Exit

All tasks checked, criteria demonstrated, /verify-live proof captured, PR opened — the Adversary loop takes it from there.