Run a quality loop

The dogfooding recipe: use the product as a demanding customer, find what falls short, fix it at the root, and leave the rubric sharper than you found it. Two entry points share the same discipline — pick by what “quality” means for this run:

• Behavior/output quality → exercising-the-product (create → observe → analyze → exercise mutations → fix → iterate; ledger in the state file).
• UI/UX quality → reviewing-ui (context brief → walkthrough → heuristic audit → prioritized batches; decision log in UI-REVIEW.md).

They feed each other: UI findings surfaced during a behavior run get logged to UI-REVIEW.md, not fixed ad hoc; behavior defects surfaced during a UI review get ledgered, and fixed via root-cause-fix.

1. Preflight — pass loop-preflight (verification automated? budget? tools?). Stack healthy, ONE dev instance, baseline gate green, browser evidence harness ready (ui-evidence). Read the existing state file / decision log first and resume, don’t restart.
2. Frame — write the acceptance contract (end state / evidence / constraints / budget) per agentic-loop into the state file. List forbidden operations explicitly (real publishes, prod writes).
3. Loop — run the chosen skill’s cycle. Evidence discipline throughout: mechanical checks before vision, re-fetch real output after every mutation, never bypass the UI, record every finding with its evidence.
4. Fix at the root — each confirmed defect goes through root-cause-fix: prove the cause, fix the class, lock it with a fail-on-revert regression test (regression-test, writing-tests).
5. Maker ≠ checker — before calling the stop condition met, dispatch the reviewer agent on the diff + contract; reconcile real findings.
6. Compound and close — append new failure modes to the rubric/matrix so the next run checks them automatically; write the closing summary (fixed / open / deliberately-left, with evidence). Edits stay unstaged; no commit/push/PR unless asked.

Stop and escalate on: budget spent while red, a fix needing a product decision, or an environment you can’t make healthy. A clean pass with zero findings is a success — never manufacture findings to justify another cycle.