AI Skill of the Day: Put Your AI Through a Builder-Critic Loop
Something Big Is Happening ● Covered by 2 sources
An AI builder shared a trick for getting way better agent output: make it fight against a real quality bar in a loop. The method turned a single Claude prompt into a 55,000-line AAA-style game, and now people are copying it for websites, writing, and more.
Matt Shumer's now-viral Call of Duty-style shooter, built from one Claude Opus 5 prompt, was so polished that plenty of people assumed it had to be faked. So he open-sourced the prompt and the code. Skeptics ran it themselves expecting it to flop. Instead they got working games, and a bunch of them started bending the same technique toward completely different projects.
The secret wasn't a magic prompt full of technical specs. It was structure. Shumer calls it a Gauntlet Loop: give an agent a goal and a real reference for what excellent looks like, let it decide how to break the job into small pieces, then assign each piece a builder and a separate critic with a fresh context window. The critic compares the work against the reference, points out the single biggest gap, and sends it back. Repeat, indefinitely, until you're satisfied or you run out of patience.
A few details matter more than they sound like they should. The agent needs a real, inspectable bar, not vague encouragement like "make it great." For the shooter, that meant actual Call of Duty screenshots. For a website, it might be the best real sites in that category; for writing, paragraphs from a writer whose clarity you admire. And the critic can never be the same agent that built the thing, because a builder grading its own work will always find its own reasoning convincing. You want a colder, blinder judgment, ideally comparing raw output — pixels, running code, finished prose — not a summary of what was attempted.
Shumer also pushes back on over-specifying the process itself. Don't hand the agent an architecture or a list of subsystems; hand it a destination and let it choose the route, including how to divide the work among subagents. He ran his own game loop for hours inside Claude Code with Opus 5 and ultracode mode switched on, letting it spawn its own fleet of workers, and he stopped only because he chose to, not because the agent hit some finish line. The game was still improving when he pulled the plug.
The method isn't limited to games. Shumer frames it as applicable to code, marketing copy, product design, or research, anything where you can hold the output up against a concrete example and ask which one is better. The pattern people are now copying is less about Claude specifically and more about refusing to accept an agent's first good-enough answer.
My take
This is basically test-driven development for creative work, and it's a little embarrassing that it took a viral shooter demo for people to notice that giving an AI a fuzzy goal produces fuzzy output. The real lesson buried in here isn't about Claude Opus 5 being magic, it's that most people are still prompting agents like they're vending machines instead of giving them a bar to clear and the autonomy to figure out how. Expect every AI tooling company to rebrand this as a paid feature within the month.
Read more about this at: Something Big Is Happening
