Free prompt tool

The Gauntlet Loop
Prompt Generator

Turn any goal into a compact prompt that makes an agent build, judge, and loop until the work beats a real quality bar.

01

Your brief

What are we making?

Why a loop beats a better prompt

Free generator that turns a goal and a named reference into a paste-ready agent prompt with a real quality bar and an independent critic.

Why it matters

An agent asked for good work will tell you it produced good work. It has no reference to fail against, so the first draft becomes the final draft. The Gauntlet Loop removes that comfort: you name something that already exists in the world, a shipped product, a published essay, a real repository, and the agent's output has to win against it.

The judging happens somewhere else. A separate critic context compares the draft and the reference without knowing which is which, so the session that wrote the draft never grades it. The loop exits on one condition only: the critic picks your work blind. Not when the agent decides it is finished.

What the tool does

This generator takes two inputs, your goal and your bar, and writes the prompt that sets all of that up: the work split into pieces small enough to judge, builder and critic kept apart, and the exit condition pinned so the run cannot end early. Paste the result into a fresh session in Claude Code, Cursor, ChatGPT, Gemini, or any agent that can reach the reference you named.

The bar is the part people get wrong. "Make it look premium" is not a bar. "Beat the Linear changelog page on typography, spacing, and motion" is: named, fetchable, and comparable line by line. A vague bar produces a vague loop.

// Gauntlet Loop Prompt OutputIteration Loop

Role: Lead Systems Builder & Independent Critic Context

Quality Bar: Linear Changelog typography, micro-interactions, and contrast ratios

Exit Condition: Blind evaluation preference over reference standard

Example generated Gauntlet Loop prompt specifying builder context, critic bar, and exit condition.

Terms worth knowing

The bar
The named reference the work is measured against. It has to be specific, reachable by the agent, and comparable on stated grounds.
Builder context
The session that produces the draft. It writes, revises, and never scores its own output.
Critic context
A separate session that inspects the real artifact against the bar. It sees the running build or the finished draft, not a summary of it.
Blind win
The exit condition. The critic prefers your work over the reference with no labels attached to either.
The loop
Build, criticize, revise, repeat. It runs until the blind win rather than until the agent sounds satisfied.

Frequently asked questions

Short answers for the decisions that usually come up after a first run.

What is the Gauntlet Loop?

A prompting pattern for work an agent grades too generously on its own. You give it a real reference to beat, a builder and a critic that never share a context, and one exit condition: the critic picks your output over the reference without being told which is which.

Is the generator free?

Yes. No account, no sign-up, no credit card, and no cap on how many prompts you generate. The prompt it writes is plain text you own.

Which AI tools does the prompt work with?

Any of them. The output is plain text with no tool-specific syntax, so it runs in Claude Code, Cursor, ChatGPT, Gemini, Copilot, or an in-house agent. The only real requirement is that the agent can reach the reference you named and see the artifact it produced.

What makes a good quality bar?

Three properties. It is named, so the agent knows exactly what to fetch. It is comparable, so the critic can hold your work beside it on stated grounds. It is currently reachable, so the comparison stays honest. A live page, a published article, a shipped game, a public repository all qualify. An adjective does not.

Why does the critic need its own context?

Authorship biases judgment. A session that just spent ten minutes writing something reads it as the author, forgives what it meant to say, and scores it kindly. A fresh context with no memory of writing it reads it the way a stranger will.

How many rounds should it run?

Until the blind win, which is usually three to five passes. If it never converges, one of two things is true: the bar is out of reach for the scope you gave it, or the piece is too large to judge in one comparison. Split the work smaller and run the loop on each part.

Does the loop work outside code?

Yes. The pattern needs an artifact a critic can inspect, not a specific medium. Landing pages, long-form writing, research briefs, design systems, and benchmark runs all fit. What breaks it is work with no observable output to compare.