← ClaudeAtlas

ai-test-looplisted

Use before declaring any work done or shipped, when the agent must verify its own output, when a deliverable needs a score before release, when the builder keeps approving its own work, or when repeated review rounds never converge. Runs an independent critic that scores the artifact 0-100 with real checks, loops until the threshold, and blocks delivery on a fail-closed security gate.
Samuelfmedeiros/ai-test-loop · ★ 1 · AI & Automation · score 74
Install: claude install-skill Samuelfmedeiros/ai-test-loop
# AI Test Loop — never ship unverified work The builder is not allowed to grade its own homework. This skill turns "looks done" into "verified, scored, and independently reviewed" — the loop either reaches the threshold with evidence, or it reports exactly which gaps remain. Use it whenever an agent produced something a human will rely on: code, UI, a pipeline, a document, a data migration. Not for throwaway exploration. ## The four roles (never merge them) | Role | Who | Job | |---|---|---| | Orchestrator | the agent running the task | picks checks, runs rounds, fixes infrastructure, owns the report | | Builder | the agent/model that wrote the artifact | fixes gaps, one class of gap per pass | | Critic | a **script you write** (`loop_runner.py` + your checks) | runs real commands, returns a 0-100 score and a gap list | | Reviewer | a **separate** agent/model context | audits critic + builder, votes, can reopen the loop | Rule zero: builder ≠ critic ≠ reviewer. If the same context does all three, the loop is theatre. Run steps 1-3 below in the working session, then drive the script for the machine part. ## The loop 1. **Declare the checks before writing code.** For each check: name, command, weight, and what "pass" means. Weights must sum to 100. 2. **Run the critic** — `python "${CLAUDE_SKILL_DIR}/scripts/loop_runner.py" --config loop.json --round N`. `${CLAUDE_SKILL_DIR}` is the directory holding this `SKILL.md`, so the path works from the project root, a su