loop-creatorlisted

# Loop Creator Builds **work loops** — a builder step, a separate verifier step, and a controlled iterate-until-it-passes cycle — plus all the prompts, queue, harness, and CI to run them. The output is a **loop package**: a spec, the prompts, a queue, a runnable harness, and (optionally) a GitHub Actions triage workflow. Everything it produces is then run through a bundled validator so the loop can't ship without acceptance criteria, a real verifier, checkpoints, and a stop condition. ## Why loops are built this way (the non-negotiables) These come straight from Anthropic's agent guidance and how the Claude Code team runs loops. They are the spine of every package this skill emits. See `references/loop-patterns.md` and `references/claude-code-loops.md` for the grounding. 1. **The doer never grades its own work.** A loop is a *generator* (builder/maker) plus a *separate-context evaluator* (verifier/checker). The verifier's job is to try to *refute* success, not rubber-stamp it. 2. **"Close is FAIL."** Acceptance criteria are explicit and testable. Partial credit is how loops drift. A verifier that can't cleanly distinguish pass from fail makes the loop circular and useless — so criteria must be concrete (a command that exits 0, a test that passes, a string that appears). 3. **Verify by evidence, not assertion.** The builder must show the command it ran and what it returned (test output, lint result, diff). Reviewing evidence is faster and safer than t