AI Test Loop
Score your agent 0-100 before you ship.
Free — v1.0.0
- Test loop with a separate critic, a 0-100 score and a shipping threshold
- Catches the "gate that never fails" and decorative criticism
- Works with any agent that can run commands
You get 6 files (12 KB) in ai-test-loop-1.0.0.zip — SKILL.md with the walkthrough plus the scripts.
sha256 5c96623d84c53970…
Stripe paymentInstant download7-day refundNo subscription
Read before buying
Your agent said “it works”. That is an opinion, not a result
Open the articleHow to start
- Declare the checks before you write codeFor each one: name, command, weight and what passing means. The weights add up to 100.
- Run the criticIt runs every command for real, keeps the raw output as evidence and prints SCORE=<0-100> and GAPS=[...]. It does not ask a model whether it looks good — it measures.
python scripts/loop_runner.py --config loop.json --round 1 - Fix only the biggest hole and run it againHit the threshold and you move to review. Missed it, take the LARGEST gap, fix it and repeat — capped at 8 rounds. If it never converges, the check set is wrong, not the code.
- Independent review before shippingSame artifact, fresh context, no justification from whoever built it — the prompt ships in the package, at references/reviewer-prompt.md.
When you will use it
- The agent grades its own homework"Looks done" is not evidence. The loop separates builder, critic, reviewer and the owner of the delivery — whoever built it does not hand out the score.
- A review round that never convergesFour reviews and nothing changes. The loop either reaches the threshold with evidence or lists exactly the gaps that remain, instead of repeating "almost there".
- A gate that never failsA check that always goes green gets called out: every verification has to be able to fail for real, and decorative criticism shows up in the report.
- Score before you shipRight before deploy you run the project's real checks and get 0-100, weighted by artifact type (code, UI, pipeline, doc).
What comes in the package
scripts/loop_runner.py — the critic: runs the weighted checks, prints the 0-100 score and the gaps, and keeps state between roundstemplates/loop.example.json — config you copy next to your project: commands, weights and thresholdreferences/scoring-rubric.md — starting weights per artifact typereferences/reviewer-prompt.md — reviewer prompt in strict JSON + the parser's anti-echo rules
Requirements
- Python 3.8+ (stdlib only)
- your project's verification commands
- any agent that can run a command
Where it does not fit: It does not deploy or merge — the decision to ship stays with the owner. It opens no network and sends no telemetry: the security step greps the repo itself.
How do I get it after paying?
The download link shows up on the thank-you page as soon as Stripe confirms, and also goes to your e-mail. If you lose it, request a resend from the same e-mail used to buy.
Can I use it on a client project?
Yes — the license covers personal and commercial use in your projects, including client work. What you cannot do is resell the package. Full terms.
Prices in Brazilian real (BRL). Stripe charges in BRL.
← all skills
Payment processed by
Stripe — no card data touches this server. Instant delivery once confirmed. Prices in BRL (Brazilian real) — you are charged in BRL.
Terms, delivery and refund.