BlogGuides

The Original Gauntlet Loop Prompt, Annotated Line by Line

Matt Shumer's Claude of Duty prompt is barely 150 words. Every line is load-bearing. Here is what each one does and what to steal.

The Original Gauntlet Loop Prompt, Annotated Line by Line

The Original Gauntlet Loop Prompt, Annotated Line by Line

The prompt that produced Claude of Duty, Matt Shumer's 55,000-line browser FPS, is three paragraphs long. It is preserved verbatim in the Wikiprompt catalog. Here is why each line earns its place.

"Build a first-person shooter at the level of the most recent Call of Duty games"

The goal names a bar in the same sentence. Not "a great FPS" - the most recent Call of Duty. That single choice powers everything downstream: the critic will need something real to compare against, and now it has one. When you adapt the prompt, this is the part to sweat. A bar must be named, fetchable and comparable; a category ("award-winning design") is none of those.

"Fan out sub-agents and have sub-agents tackle each one individually"

One agent building everything produces mush. Splitting into narrow pieces (render, physics, weapons, audio...) means each builder produces a real, checkable artifact. Claude of Duty ended up with 11 subsystems, each owned separately, coordinated by an architecture contract.

"Have a separate sub-agent check it visually... a really harsh critic"

The load-bearing line. The critic is SEPARATE: fresh context, no memory of the builder's effort, no incentive to be kind. "Really harsh" matters too - a polite critic converges on approval. The critic's job is not encouragement; it is judgment.

"It should literally compare them side by side blind and say which one looks better"

No scores. Scores out of 10 inflate round after round. A blind side-by-side pick cannot inflate: either the work beats the bar or it does not. This is the mechanism that keeps the loop honest for hours without a human in it.

"Don't stop until each sub-agent is utterly wowed... /loop until it's utterly perfect"

The exit condition is winning, not a round count. Fixed iteration counts ("improve it 5 times") stop exactly when the model feels like coasting. The gauntlet exits when the blind comparison flips.

What to steal

Steal the structure, not the words: goal with a named bar, builders on narrow pieces, a separate harsh critic, a blind binary comparison, and an exit that is earned. The reusable template packages exactly that. More angles in What is the Gauntlet Loop and failure modes.

Tags
gauntlet-loop·agent-loop·prompt-engineering·claude-code·builder-critic·agents