What is gauntlet-loop?
gauntlet-loop is an open-source agent skill that writes a short prompt for a goal-driven evaluation loop. The skill serves agent users working on websites, essays, command-line tools and research briefs. Rather than completing the task directly, gauntlet-loop produces instructions for another agent session to build, critique and revise the work against an existing reference.
gauntlet-loop packages a technique created and named by Matt Shumer while building Claude of Duty. The repository credits Jay E at RoboNuggets for the skill. The distinction matters: the repository supplies a reusable prompt-writing skill, while the underlying technique supplies the builder-critic structure, blind comparison and continuation rule.
What gauntlet-loop can do
gauntlet-loop turns a broad goal into a comparison-based workflow anchored to a concrete quality bar. Before writing the prompt, gauntlet-loop checks that the reference is named, fetchable and comparable. A reference can be a published page, a post or a repository, provided the critic can inspect the reference and judge the output alongside that reference.
- gauntlet-loop offers 2 or 3 quality bars for the user to choose from.
- gauntlet-loop writes a paste-ready prompt of around 150 words.
- gauntlet-loop instructs the agent to divide work into smaller pieces.
- gauntlet-loop assigns builders and separate critics with fresh context.
- gauntlet-loop uses label-stripped comparisons rather than numerical grading.
- gauntlet-loop continues until the output wins or the user stops the run.
How gauntlet-loop works
gauntlet-loop is installed by cloning the GitHub repository and copying the skill folder into a project's .claude/skills directory. The user invokes /gauntlet-loop with a goal, selects a proposed quality bar and receives a prompt. The user then pastes that prompt into a fresh session, where the agent performs the iterative work.
gauntlet-loop uses Claude Code's /loop and ultracode features for repeated execution and multi-agent orchestration. For other agents, gauntlet-loop replaces those lines with plain instructions to continue until the critic selects the generated work and to run builders and critics as parallel subagents. The repository describes the workflow structure as identical across these alternatives.
Who gauntlet-loop is for
gauntlet-loop is for people who use agents to create work that can be compared with an accessible reference. The repository's examples cover visual design, explanatory writing and software development. Each example ties the task to an inspectable standard, such as a live campaign page, published writing or a tool implementation with a benchmark.
- gauntlet-loop can guide landing-page work against desktop and mobile screenshots.
- gauntlet-loop can compare an explainer with a named writer's published posts.
- gauntlet-loop can guide a JSON log formatter against an implementation and benchmark.
- gauntlet-loop can generate a workflow prompt for a research brief.
gauntlet-loop pricing and licence
gauntlet-loop is free to use with attribution under CC BY 4.0. The repository includes the skill in a SKILL.md file, alongside a README and licence. Access is through the public GitHub repository, and the documented setup consists of cloning the project and copying the skill folder into the user's project.
gauntlet-loop strengths and limitations
gauntlet-loop separates creation from evaluation and grounds criticism in existing work rather than a self-authored rubric. The repository also identifies conditions that break the process: vague references, builders judging their own work, soft critics and fixed round counts. The workflow depends on an independent critic that can actually access and inspect the chosen reference.
- gauntlet-loop requires a named, accessible reference rather than a vague quality target.
- gauntlet-loop requires fresh critic context separate from the builder.
- gauntlet-loop does not stop after a fixed number of rounds.
gauntlet-loop FAQ
What is gauntlet-loop?
gauntlet-loop is a reusable agent skill that turns a goal into a short prompt for iterative building and criticism. The generated prompt directs an agent to compare its output blindly against a concrete reference and keep revising until the output wins or the user stops.
Is gauntlet-loop free?
gauntlet-loop is free to use with attribution under CC BY 4.0. The public GitHub repository contains the skill file, README and licence, and documents installation by cloning the repository and copying the skill folder into a project.
How do you install and use gauntlet-loop?
gauntlet-loop is installed by cloning the repository and copying its skill folder into the project's .claude/skills directory. The user invokes /gauntlet-loop with a goal, chooses a reference quality bar, and pastes the generated prompt into a fresh agent session.
Does gauntlet-loop work outside Claude Code?
gauntlet-loop supports other agents by replacing Claude Code's /loop and ultracode lines with plain instructions. Those instructions tell the agent to keep looping until the critic chooses the generated work and to run builders and critics as parallel subagents.
What can you use gauntlet-loop for?
gauntlet-loop can generate workflow prompts for websites, essays, CLI tools and research briefs. The repository gives examples involving a brand landing page, an explainer on vector databases and a JSON log formatter, each evaluated against a specific accessible reference.
When does a gauntlet-loop run stop?
gauntlet-loop instructs the agent to stop when the generated work wins a blind comparison against the selected reference, or when the user stops the run. The documented technique explicitly rejects ending after a fixed number of rounds.
From the community
No reviews yet
Be the first to review this tool. What you write helps the next person choose.




