Give your agent this
Paste this into Claude Code or Grok Build, or any assistant that can browse.
Read https://ratchet.babzlabz.com/agents.md and help me fill in the Ratchet Loop Generator for this project: <describe your project>
If it cannot browse the web, paste the contents of
/agents.md into the conversation instead.
This tool supports two: Claude Code and Grok Build. That is a choice, not a gap. A run kit installs a boundary hook, names the command that starts a run and writes a context file in that harness's own format — so a kit for a tool nobody here has run would be a set of instructions that do not work.
Choose the harness separately from the model
Your coding model can change without changing the method. Choose Grok Build or Claude Code for the window that will run the kit, then select a model that that installation actually offers. A display name is not necessarily a valid command-line model identifier. Keep the model and effort setting fixed within a comparison, and record changes.
The builder and its judges have different jobs. Give judges fresh contexts, no builder summaries, and the most independent models the harness supports. Verify the resolved model from the harness where available; a model naming itself is not proof. Where the harness offers models from three or more vendors, both judges come from vendors other than the builder’s. With one vendor the panel is recorded as single-vendor: a limitation to disclose, not four independent opinions.
What it tells them
The document is written for a machine, so it is blunt and specific. It covers:
- All six fields — what each needs, what format, and which three are required.
- Which window, not which model. Field 03 is the CLI the prompt will be pasted into. A kit made for Claude Code carries a hook and role files Grok Build never reads, and the reverse; every real run that was generated for one and pasted into the other ran with no boundary at all. The agent is told to ask which window, not which AI.
- The formatting that changes behaviour. More than about three bullet points in the goal makes the tool pick the most expensive tier, which is right for a big multi-part project and wasteful for a small one. A round count picks no tier.
- Reference rules learned the hard way. Direct image files beat gallery pages, because galleries lazy-load and screenshot as blank thumbnails. Some sites serve a bot check instead of the page, which screenshots perfectly and is worthless. A gallery folder often returns “not found” while the images inside it work fine.
- How to write the one check that is yours. One property per entry, never two joined with “and”, and something binary beats something subjective.
- Budget units and their limits. Whole tokens on Grok Build, estimated USD on Claude Code. Boundary estimates can overshoot; a missing reading is not free work.
- An output format, so what comes back is ready to paste rather than an essay.
An ordinary run or a supervised beta run
The ordinary kits support both Grok Build and Claude Code. Follow
run-kit/START-HERE.md, merge the project settings rather than replacing your own,
run the kit’s collision-checking installer, prove the hook works, and use the
kit’s hand-off. Differing local hooks or roles stop installation; they are never
overwritten automatically.
Relay 5.0-beta.3 is an optional Grok supervisor for Duel or Ratchet.
It starts a fresh headless session per round, checks the record, and applies the configured
limits. It runs with approvals off. Start with its doctor and a bounded
run --once; synthetic tests are not evidence of a successful unattended paid run.
Keep an existing lineage on the kit and method version it started with.
Why this exists
The generator has six fields, three of them required, which makes it easy to fill in badly. The expensive mistakes are not obvious ones: a kit generated for one CLI and pasted into the other, a reference that looks fine but returns a bot-check page, an invariant that quietly tests two things at once, a bulleted goal that picks the most expensive tier for a project that is really one piece.
Those are exactly the mistakes an assistant can catch for you, if it knows the rules. So the rules are written down in one place, in a form a machine can read.
One thing to tell it about references
An agent that cannot reach the web will happily invent plausible-looking URLs. Ask it to say which references it actually verified, and treat the rest as guesses. The generator will now tell you when a capture came back as a bot check or a missing page — the reference shows as “not the page” rather than quietly succeeding — but it is cheaper to catch that before you start a run than after.
Keep secrets out of uploads
Both pickers put a file at a public URL: an uploaded reference when you press Upload, a check file as soon as it is selected. Anyone with the link can read it for about 12 hours. “Sealed” means the builder should not change the check; it does not mean the online copy is private. Use a sentence or a secret-free check, and download the kit while the uploaded copy is available.
Write a brief that can settle whether this works
Name the observable outcome, the state in which it matters, the reference to compare, and a check that would fail on a real defect. Ask for evidence from the running artifact, not a summary of the code. A claimed improvement should survive a fresh checkout and another person repeating the interaction.
Ratchet’s goal is better delivered work for the resources spent. Extra rounds, extra judges, or a long-running process do not prove that. Compare against Gauntlet with the same starting point, model, bar, and budget, then judge the delivered versions blind. Read the comparison standard.