Give critics something concrete.
A design system, test suite, acceptance criteria or reference build makes disagreement useful. Without one, the loop can polish the wrong direction.
← All tools
Open-source Agent Skill
Good work shouldn’t survive on self-belief.
Gauntlet Loop turns your brief into a contest of evidence: specialists make it, critics try to break it, and revision continues until the work clears a bar you set.
- Claude Code
- Codex
- tfcode
01BuildSpecialists fan out
02RefuteCritics hunt for proof
03ReviseWorkers close the gaps
One skill.
Separate workers and critics.
A finite quality bar.
How it works
Make the work earn “done”.
The skill turns your task into a compact orchestration prompt with three things agents usually leave vague: who owns each piece, who challenges it, and what evidence ends the loop.
-
01
State the task
Define the thing to build, the real constraints and the reference that represents quality.
-
02
Fan out the work
Independent specialists own bounded pieces. Each gets a separate critic whose job is to refute, not politely assess.
-
03
Close the loop
Workers revise against concrete failures. The loop stops at the quality bar—or at the pass cap, with unresolved issues surfaced.
gauntlet-loopSKILL.md
TASK
Build the thing. Define done.
BUILD METHOD
Fan out specialists. Pair each
with a critic that must refute.
BAR
Revise until evidence clears it.
Cap every loop.Use it in three moves
- 1
Install the skill in your agent’s skills directory.
- 2
Ask for a gauntlet loop and describe what you are building.
- 3
Run the prompt in Claude Code, Codex or tfcode.
Use the horsepower wisely
A gauntlet is powerful. It is not cheap.
Parallel workers, independent critics and revision passes multiply token use and elapsed time. Reserve the full loop for work where another pass is worth the spend.
Expect more tokens and time.
Fan-out and retries can run for an hour or more on a substantial build. Use a strong MVP as the starting point, then spend the loop on the finish.
Always cap the passes.
“Make it great” has no natural endpoint. Set a maximum per piece and surface anything that still fails instead of allowing an infinite loop.
Origins
Gauntlet Loop was inspired by Matt Shumer’s “Claude of Duty” experiment in adversarial, multi-agent build loops. This repository is an independent, portable Agent Skill implementation with no affiliation with, or endorsement by, Matt Shumer or Anthropic.
Ready for the critics?
Put your next important build through the gauntlet.
One open-source Agent Skill. Three agent environments. A much harder path to “looks good to me”.