The method
Make it hard to pass, and the work gets good
Agents are already good at building. What they are bad at is knowing when they are done. The Gauntlet method fixes the second problem and refuses to interfere with the first: give the agent a goal and a bar, then put someone in the room who did not do the work and whose only job is to find what is wrong with it.
Gauntlet Loop
The original method
A lead agent splits the goal into pieces it can judge independently. Each piece gets a builder and a separate fresh-context critic. The critic compares the real output against the bar and sends back the single biggest gap. Repeat until you stop it.
Goal, not implementation
Every sentence about how to build it is a sentence the agent no longer has to think about. The bar stays; the plan is the agent's problem.
A concrete, inspectable bar
A critic can hold work up against Linear's onboarding or a test suite. It cannot hold work up against "high quality".
The lead agent decomposes
You do not know where the independent seams are until you have looked at the problem. The agent has; you have not.
The builder never grades itself
An agent asked to check its own work grades the plan it remembers, not the artifact it produced.
Critics get fresh context
A critic that watched the work being made inherits its excuses. One that arrives cold sees what a user would see.
Critics inspect the real artifact
Running the app, reading the file, looking at the render. A summary of the work always sounds better than the work.
No fixed round count
A number is a finish line, and a finish line is what stops the loop before the bar is met. It runs until you stop it.
A live progress page
The run is long and mostly invisible. One always-current page is what makes stopping at the right moment possible.
The Gauntlet Loop is Matt Shumer's method, and the credit for it is his. Read the original write-up. This site is an independent tool that generates prompts implementing it, plus our own extension below. Nothing here is copied from it.
Gauntlet Graph — our extension
Replace the critic with a gauntlet
A single critic per piece has one blind spot: whatever it is not looking at. Graph mode replaces it with a directed graph of adversarial trials, each run by a separate fresh-context critic attacking one distinct quality, and makes surviving the graph the only definition of done.
A challenge graph per piece
One critic tests one thing well. A graph of them tests a piece the way reality will — from several directions, none of them coordinated.
One dimension per trial
A critic told to check five qualities checks the easiest one properly and the rest in passing.
Clean single-pass traversal, or nothing
Partial credit is how mediocre work ships. Four of five trials passed is a failure, and the piece knows it.
Failure resets to the entrance
A fix that breaks an earlier trial has to be caught, and the only reliable way to catch it is to run the earlier trials again.
The graph engineer adds a node
Anything that survives a fixed test suite eventually games it. A gauntlet that grows a new trial after every survival cannot be gamed for long.
The graph never softens
Nodes are added, never removed. The bar can only ratchet upward, which is the whole point.
One line about what you want. That's the input.
Three prompts a day, free. The generator does the method; you supply the ambition.
Write my prompt