Free · 3 prompts a day
Prompts that make your agent earn it
Describe what you want built. You get back a short, paste-ready prompt that hands your coding agent a goal, a bar, and a set of adversarial critics it has to get past — and nothing about how to build it.
Gauntlet Loop
A builder and a separate fresh-context critic per piece. The critic compares the real output against the bar and sends it back until it clears.
Gauntlet GraphOur extension
A whole graph of adversarial trials per piece. One failure resets the piece to the entrance — and every survival adds a harder trial.
Free — no account
One email, then the generator is yours
Three prompts a day, every day, for nothing. In exchange you get the occasional email about the method — what people are building with it, and which trial patterns actually catch things.
What happens next
The prompt is short because the method is strict
Everything the generator writes is a rule about who judges what, and how hard. Nothing in it is a plan — that is the part your agent is actually good at.
01
You give it a goal
One line about what should exist. Not a spec, not a stack, not a file layout — the method breaks the moment you start prescribing implementation.
02
It writes a sparse prompt
Under 350 words, plain text, paste-ready. It carries the goal, the bar and the rules of the method, and deliberately says nothing about architecture or how many rounds to run.
03
Your agent decomposes and builds
The lead agent splits the goal into pieces it can judge independently, then hands each to a builder subagent. It chooses the split; the prompt never does.
04
Fresh critics try to break it
Each critic starts with no context, looks at the real artifact rather than a description of it, and sends back the single biggest gap. The builder never grades its own work.
Graph mode
One critic is a check. A graph of them is a gauntlet.
In Graph mode each piece faces a directed graph of trials, one fresh critic per quality dimension. Passing four of five is not a pass — any failure sends the biggest gap back to the builder and puts the piece back at the entrance. Survive a clean run and a graph engineer adds a harder node. It only ever gets harder.
Before you run it
The parts people ask about twice
Mostly because the method's best ideas look like omissions the first time you read them.
Why is the generated prompt so short?
Because everything it leaves out is a decision your agent is better placed to make. The method's whole claim is that a goal plus a bar plus a critic who cannot be fooled beats a detailed specification. Adding paragraphs about architecture measurably makes the output worse.
Why doesn't it say how many rounds to run?
A round count is a finish line, and a finish line is exactly what the method removes. The loop runs until you stop it — which means you are judging the artifact rather than watching a progress bar fill.
What does a run actually cost?
More than you expect. Every piece is built at least twice and judged by a fresh critic each time, and in Graph mode a late failure re-runs every trial before it. Watch your budget for the whole run. This tool writes prompts; what they spend is on you.
Do I need an account?
No. An email address unlocks the generator and gets you three prompts a day. You can unsubscribe from any message with one click, and your prompts stay readable to you on this browser.
Is this the Gauntlet Loop from somethingbig.ai?
The Loop is Matt Shumer's method and it is credited on the method page. Gauntlet Graph is our extension of it — the challenge graph, the reset-on-failure rule and the graph engineer are ours.
Still curious? Read the method in full.