Prompt generator Claude Code · Grok Build No gate, no limit, no truncation

Ratchet
Loop

A build loop where quality can only click forward. Say what you want built and what good looks like — this writes the prompt, captures your references, and hands back the run kit. Every round has to beat the round before it, not just the reference.

What is a Ratchet Loop

0 / 6 · advances only

Brief No — Spec sheet · every field reachable, none required but one
0 chars

Say what the finished thing has to achieve. Don’t prescribe the stack or the architecture — that caps the result at your own imagination. There is no length limit and nothing here is ever truncated; the field scrolls rather than pushing the page.

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

A bar has to be named, fetchable and comparable — something the agent can open, run, read or screenshot and put beside its own work. Leave it empty and the prompt makes finding one round zero’s only job.

One surface, three kinds. Drop a file here — or use any of these, all keyboard-reachable:

Typed path
The agent opens it directly — only when it runs on this machine.
Dropped file
The browser hands over the name only, never the folder. Give the folder once; it is kept for next time.
Uploaded
Becomes an https URL any agent can fetch, anywhere. Use this for anything shared with a team.
A URL
Captured here — desktop and mobile full-page screenshots plus the page as markdown. Real pixels, or a labelled blank.
  • Capture · Kind · Reference · StateThe bar is empty. Every reference you add rules a notch across this box — the literal string the agent is handed, whichever of the three tests it fails, and real pixels where a capture exists.
03 · Where it runs Claude Code
04 · Loop strength Gauntlet

Gauntlet by default on an empty brief; this line always says why, and picking a plate overrides it.

05 · Caps — optional, and you should no caps

Cost This family of methods routinely runs into hundreds of dollars. A promoting Ratchet round costs roughly 5–9× a plain Gauntlet round. Nothing here stops it but you — set a cap you are comfortable losing, and watch it.

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

Every other check comes from the agent reading your words. This one does not. It runs every round and the agent may not edit it, so it is the only line in the whole run that is yours. One sentence is enough.

Plain words. The agent builds whatever harness demonstrates it — your sentence is fixed and the harness has to fit it, never the other way round.

Nothing uploaded. If you have a script, it is the strongest version of this: nothing about it gets interpreted.

Already included, whatever you type here. It must not log errors, must behave the same way twice, must be reachable by keyboard, must not call out to anyone else’s server, and must never get slower or heavier than the first round made it. Those last two carry no number: round 0 measures what it built and that becomes the ceiling. The agent writes down how to start the thing once it exists, and where a check genuinely cannot apply it has to say so in writing — which you can then disagree with.

The second one writes a different, much shorter prompt: go and find three to five real references, propose the criteria with the measurement that decides each, propose the sealed suite, sketch what it would and would not build — then stop, so you can disagree before the money is spent. Cut what is wrong and bring it back here.

The filed brief

Unfiled 


        
        
Nothing filed yet.


          The prompt lands here — complete, self-contained, and never truncated at any length.
          It carries the invocation, the goal, the bar with every reference and how to obtain each,
          the loop mechanics, the harness primitives, the caps and the stop conditions.