01 / The generatorMethod 4.5

Ratchet
Loop

Build. Compare. Keep the better version.

Give Grok Build or Claude Code a goal, a real reference, and a check that must hold. Get a complete run kit. Keep the strongest accepted version, not simply the latest change.

Fig. 01 / The championAccept · preserve
Progress is what you keep.A new attempt must earn its place against the accepted version.
  1. Build ×2
  2. Blind duel vs the champion
  3. Win → keep
  4. Lose → preserve champion

Claude Code · Grok Build
No account. Complete prompts.

Brief No — 01, 02 and 06 are required
0 chars

Say what the finished thing has to do, not which tools to use.

Why, and what works

Naming the tools limits the result to what you already thought of. Write as much as you like: nothing is ever cut, and the box scrolls instead of pushing the page down.

02 · The bar — what does good look like? no references

Something real the agent can open beside its work: a URL, a file in your repo, or a sentence. Nothing yet? Use I don’t know the bar yet at the end of the form.

Pick one of these, or drop a file on this box:

Uploading makes the file readable by anyone with its link for about 12 hours. Keep private material out.

Which references work
Typed path
The agent opens it directly. Prefer one relative to your repo — refs/ or ./refs/hud.png — because it resolves wherever the run works. An absolute path like C:\refs opens only on this machine.
Chosen file
Browsers give us the file name, never the folder. Tell us the folder once and we remember it.
Uploaded
Becomes a public web link the run can open from anywhere, for about 12 hours; the kit keeps its own copy.
A web link
Screenshotted for you, desktop and phone, plus the page text. A bot check or an error page is flagged not the page.
  • Capture · Kind · Reference · StateThe bar is empty. Everything you add shows up here: what the agent will be given, anything wrong with it, and the screenshots where we have them.
03 · Where it runs — the window you paste into, not the model Claude Code
04 · Loop strength — what differs is attempts a round, and what you ship Duel

Cost is per round, against one Gauntlet round: estimated from what each round spawns, not measured.

We pick one for you from your brief. This line says why; clicking a plate overrides it.

05 · Caps — optional, and you should no caps

USD, from Claude Code’s estimated session cost. This is a prompt-managed boundary limit, not a hard billing ceiling.

Whole rounds. A number here is a limit, never a choice of loop.

Blank means no limit, and the prompt says so in words. Runs of this shape can reach hundreds of dollars.

What a cap really does

Cost Nothing here stops spending but you — set a cap you can afford to lose, keep an eye on it, and keep a provider-side spending limit as well.

A limit does not squeeze the work into the budget. The run goes at full speed, stops before a round it cannot afford to finish, and hands you the best version it reached with a note saying it ran out: fewer improvements, not a half-built thing.

06 · What must always be true — required, and worth more than the rest nothing of your own

One sentence: something someone does, and what proves it happened.

How to write one

Everything else the agent checks, it decided on by reading your brief. This one it cannot touch or reword: it runs every round exactly as you wrote it. Keep it to one property — two joined with “and” cannot tell you which half failed.

Every kit also carries checks for errors, repeatability, keyboard access, and unexpected third-party calls. The binding checks can reject a candidate in Duel and Ratchet. Speed and transfer size are measured against round zero but are advisory. Gauntlet carries findings to its critic rather than using a gate. A check that cannot run is reported as unrunnable, never as a pass.

Selecting a file uploads it to a public URL. Anyone with the link can read it for about 12 hours, so keep credentials and private data out. “Sealed” means protected from the builder, not private online.

Nothing uploaded. A Node.js check must print a final JSON line with satisfied: true or false and exit 0.

Captures Nothing pending

Web links on the bar are screenshotted before you file — desktop, phone and the page text. Nothing is waiting on one.

Cannot file yet — 01 · The goal, 02 · The bar, 06 · What must always be true still empty. “I don’t know the bar yet” needs only the goal.

Filing needs 01, 02 and 06. The second button writes a much shorter prompt: find three to five real references, propose the criteria and the sealed check, then stop before anything is built, so you can disagree before the money is spent.

The filed brief

Unfiled


        
        

Nothing filed yet.

Your prompt appears here in full — never cut short. The kit saves it as run-kit/PROMPT.md, which is what the run reads: your goal, every reference and how to get it, how the loop works, your limits, and when to stop.

02 / The method

Why keep a champion?

Understand the loop, the evidence, and the limits.

03 / For agents

A better brief starts here.

Give your assistant the field guide, not guesswork.

04 / Relay beta

A run beyond one session.

Explore supervised rounds for Grok Build.