Gauntlet Loop — core skill (domain-agnostic)

A disciplined loop for producing high-quality work in any domain: build → blind critique → revise → repeat, against a bar the agent cannot talk its way past. Coding, writing, design, data, research, prompt engineering and detection all use the same loop — they differ only in how you pick the bar, what the critic inspects, and the checklist. Those live in references/domains/.

Origin: the Gauntlet Loop method is Matt Shumer's, from "How to Run a Gauntlet Loop" (https://somethingbig.ai/gauntlet-loop), with the companion repo mshumer/Claude-of-Duty. This skill is an independent, attributed adaptation — see ../../CREDITS.md.

Prerequisite: a real agent harness

This loop requires an agentic harness that can open files, run code, render output, take screenshots, call tools, and spawn sub-agents with their own clean context (Claude Code, Cowork, Codex, or equivalent). A single-turn chat cannot produce an independent blind critic — if you can only reply in one context, say so and fall back to a single self-review pass instead of pretending the loop ran.

For large multi-agent runs, use the harness's parallel-agent facilities (in Claude Code: sub-agents, or /ultracode when available).

The four pillars

A bar the agent cannot argue around. The strongest form: the artifact must match or beat something real — a reference implementation, a test suite, a top product's screenshot, a model essay, an eval set, an ATT&CK technique, a working PoC. The bar may be aspirational: it does not have to be realistically reachable. An unreachable bar keeps the loop pulling upward instead of stopping at "good enough".
Give the goal, not the implementation. State what the finished thing must achieve; let the agent choose the route. Prescribing architecture replaces the model's judgment with yours and caps the result at your imagination.
Let the agent split the work. The lead breaks the goal into the smallest pieces that can be improved and graded independently. Independent pieces can run as parallel loops.
The builder never grades itself. The builder and the judge are different agents with separate context. The critic is blind to the builder's reasoning and inspects the real artifact — running code, rendered pixels, actual test output — never the builder's summary.