Gauntlet Loop skill
A skill that turns a goal into a ready prompt: the agent works in builder and critic loops until its output wins a blind comparison with a reference
- 1,167 ★ on the repository
- Last commit
- CC-BY-4.0
Minimal access
Why this level
- After «run it» the agent starts a loop with parallel subagents and token use grows
- The critic opens real pages and repositories from the bar and reads their content
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/gauntlet-loopDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- After «run it» the agent starts a loop with parallel subagents and token use grows
- The critic opens real pages and repositories from the bar and reads their content
This is third-party code. Review the repository files before installing.
What it does
Gauntlet Loop writes a prompt for another session and does not deliver the finished work itself. You name a goal, such as a landing page or an article. The skill offers two or three quality bars, meaning concrete references the agent can fetch and compare against: a known product's page, a named author's published piece, a repo with a benchmark. After you pick, it returns one block of about 150 words for a fresh session. With that block the agent splits the work into small pieces, runs a builder and a separate harsh critic with fresh context on each, and the critic blind-compares the work with the reference with labels stripped and names the biggest gap. The loop runs until the critic picks your work. Matt Shumer devised the technique, and this repository packages it as a skill.
Who it is for. For people who run long tasks in Claude Code and want the agent to check against a real reference rather than its own judgment.
Good fit when
- You need a landing page for a brand, and the bar should be the live page of a known product, screenshotted on desktop and mobile
- You need a 2000-word explainer for non-specialists, with the critic comparing it to published posts by a named author
- You need a tool that formats JSON logs, with a bar that combines another repo's implementation and its benchmark
- The agent stops at «good enough» and you want an exit on winning a blind comparison, not on a round count
Not a fit when
- You need the result right now: the skill writes a prompt, and a fresh session or this same agent after you say «run it» does the work
- You cannot name a concrete bar: a vague bar is, in the author's words, the most common failure, and the critic approves everything
- The task is small and fits one pass: a loop of agent pairs is overkill for it
Example request
/gauntlet-loop a landing page for my running brand, dark and green, has to feel aliveLimitations
The /loop and ultracode commands belong to Claude Code. For other agents the skill replaces the last two prompt lines with a request to loop until the critic picks your work and to run builders and critics as parallel subagents. If the reference cannot be fetched (a screenshot, a text, a repository), the critic invents a comparison. The skill adds a budget line only if you named one. The CC BY 4.0 license requires attribution.
How to disable. Delete the .claude/skills/gauntlet-loop folder from the project. Prompts it already produced remain plain text.
FAQ
How do I install Gauntlet Loop in Claude Code?
Clone the repository and copy the .claude/skills/gauntlet-loop folder into your project's .claude/skills/. Then call /gauntlet-loop with your goal. The whole skill is a single SKILL.md file.
What is a quality bar in a gauntlet loop?
A reference that passes three tests: it is named specifically, the critic can fetch it (screenshot, read, run) and it can sit next to your work. «Award-winning sites» fails, the page of a named product passes.
When does the loop end?
When the critic blind-picks your work or when you stop the run yourself. There is no fixed number of rounds.
Why does the critic pick a winner instead of giving a score?
According to the author, scores out of ten drift upward round after round. A binary choice of which work is better does not drift.
Whose technique is this?
Matt Shumer named it and wrote the original prompt while building Claude of Duty. The RoboNuggets repository only packages the idea as a skill.
Discussion
No questions or reviews yet.
Related
SkillsA skill for creating skills that proves them by behavior: before-and-after runs, regression evals and ecosystem audit
A skill that designs, tests, evaluates and packages other agent skills, starting from architecture
A skill that scores a prompt on five criteria, picks one of 31 frameworks and rewrites the request after clarifying questions
A completion-discipline skill for large agent tasks: write the acceptance ledger with runnable checks first, then do the work, then reverify against evidence
A Claude Code skill that takes a founder from idea to a live Claude Managed Agent: interview, v0 build, launch, grading and scheduling
A parallel divergent ideation skill for agents: fans out under cognitive frames, prunes traps and deepens the best branches
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| Gauntlet Loop (this entry) | no | 1,167 | ||
| dream-loop | no | 1,747 | a loop skill for scenes where the agent generates its own reference screenshot and iterates through a critic, while here the user picks the reference | |
| Ralph Orchestrator | no | not required | 3,190 | a CLI that loops agents until a task is done, with no blind comparison against an external reference |
README in short
The README presents the skill as a one-prompt generator and explains how a bar differs from a rubric: a rubric makes the agent grade itself in its own words, a bar compares it with an existing reference. Then come a quick start, the repository contents (one SKILL.md file and a license), four steps of operation, three example goals with bars, a note on porting to other agents, a list of ways the loop breaks and credit to Matt Shumer.
README badge
Original SKILL.md
--- name: gauntlet-loop description: Turns any goal into one short, paste-ready "gauntlet loop" prompt - a prompt that makes an agent set a concrete quality bar, split the work into small judgeable pieces, run a builder and a separate harsh critic on each, compare blind against the bar, and loop until it wins. Works for builds, writing, code, research, or design. Triggers on "/gauntlet-loop", "gauntlet loop", "gauntlet this", "make a gauntlet prompt", "loop until it beats X". --- # Gauntlet Loop The user gives a goal. You give back ONE short prompt they can paste into a fresh agent session. You are not doing the work. You are writing the prompt that makes another agent grind on the work u