01 · The goal — what do you want built?
0 chars
Say what the finished thing has to do . Don’t name the
tools or how to build it — that 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
Reference kind
Auto
URL
Path
Words
Reference — a URL, a path on disk, or words
Add
Give it something real to aim at — a site, a file, a screenshot, or just a sentence. It has to be something the agent can actually open and hold up next to its own work. Leave this empty and finding one becomes the run's first job.
Drop a file here, or pick any of these — they all work from the keyboard too:
Paste a URL
Type a path
Choose a file
Describe it in words
Upload it → https URL
Typed path The agent opens it directly. Prefer one relative to your repo
— refs/ or ./refs/hud.png — because that resolves wherever the
run works. An absolute path like C:\refs only opens if the agent is on this machine, and
a sandboxed one is not.
Dropped file Browsers only give us the file name, never the folder . Tell us the folder once and we remember it.
Uploaded Becomes a web link any agent can open from anywhere . Use this if you are sharing with a team.
A web link We screenshot it for you, desktop and phone, plus the text of the page. Real pictures, or an honest blank if it would not load.
Capture · Kind · Reference · State The bar is empty.
Everything you add shows up here — exactly what the agent will be given, anything wrong with it, and the screenshots where we have them.
03 · Where it runs
Claude Code
Claude Code
/effort max then paste the brief
Setup: one clean session — no unrelated skills or MCP servers.
Grok Build
/goal <objective> --budget N one line, kit on disk
Setup: turn goal mode on, then trust the folder or .grok/hooks/ is silently skipped.
04 · Loop strength — each one adds to the one before it
Duel
Gauntlet
1× cost
build → get judged → fix repeat
The original. Cheapest. Hands you the last thing it built, and nothing notices if a later round breaks an earlier fix.
Duel
~3× cost
+ every round must beat the best one + checks the corners nobody tried
Hands you the best round instead of the last. Good for most jobs, which is why it is the default.
Ratchet
~12× cost
+ re-tests everything that passed every single round
Nothing that worked can quietly stop working. Worth it on long jobs with many parts — not before.
We pick one for you from your brief. This line always says why, and clicking a plate overrides it.
05 · Caps — optional, and you should
no caps
Leave one blank and the prompt says so in words , so the run knows you
meant it and cannot invent a limit of its own. That has bitten someone: they left this empty
expecting no limit, the run read “set a maximum round count”, picked 8, and stopped
there. A Ratchet run wants 25–30 rounds to be worth its price.
Cost These runs can reach
hundreds of dollars . Nothing here stops that but you — set a cap you can afford
to lose, and keep an eye on it.
06 · What must always be true — optional, and worth more than the rest
nothing of your own
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, which
makes it the only part of the run that is entirely yours . One sentence is enough.
One thing that must never break — and what proves it
Plain words. The agent works out how to test it — your sentence
stays exactly as written, and the test has to fit it, never the other way round.
These are always checked, whatever you write above. No errors
in the console, the same result twice in a row, everything reachable by keyboard, no calls
out to anyone else’s servers, and never slower or heavier than the first version was.
Those last two have no fixed number — the first round measures itself and that becomes
the limit. If a check genuinely does not apply, the agent has to say so in writing, and you
can disagree.
File the brief
I don’t know the bar yet
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.