Gauntlet Loop

A skill that turns any goal into one short, paste-ready prompt. That prompt makes the agent pick a real quality bar, split the work, run a builder and a separate harsh critic on each piece, compare blind against the bar, and keep looping until it wins.

This repo is the marketplace for five harnesses. Each one gets its own spawn and loop primitives — Claude Code uses /loop and ultracode; the others do not.

Adapted from robonuggets/gauntlet-loop. The technique is Matt Shumer's.

A Grok-only fork lives at aaronwJordan/gauntlet-loop-grok. This repo is the portable one.

Install

This repository is the marketplace. Add aaronwJordan/gauntlet-loop, then install the gauntlet-loop plugin.

Harness	Add marketplace	Install
Claude Code	/plugin marketplace add aaronwJordan/gauntlet-loop	/plugin install gauntlet-loop@gauntlet-loop
Codex CLI	codex plugin marketplace add aaronwJordan/gauntlet-loop	codex plugin add gauntlet-loop@gauntlet-loop
Copilot CLI	copilot plugin marketplace add aaronwJordan/gauntlet-loop	copilot plugin install gauntlet-loop@gauntlet-loop
Grok Build	grok plugin marketplace add aaronwJordan/gauntlet-loop	grok plugin install gauntlet-loop@gauntlet-loop --trust
OpenCode	clone, then point skills at this repo's skills/ directory	see below

Grok requires --trust to install, and grok plugin marketplace update gauntlet-loop before an install picks up a new release. Installing the repo URL directly instead of through the marketplace also works, but Grok then names the plugin after the URL, so its agents come out as <hash>:gauntlet-builder rather than the gauntlet-loop: prefix the skill's run files assume.

Then:

/gauntlet-loop build me a pricing page for my SaaS

It offers 2 or 3 bars, you pick one, it writes one prompt. One line under that: it can run the loop here.

OpenCode (any project)

git clone https://github.com/aaronwJordan/gauntlet-loop.git ~/.config/opencode/vendor/gauntlet-loop

{
  "$schema": "https://opencode.ai/config.json",
  "skills": ["~/.config/opencode/vendor/gauntlet-loop/skills"]
}

Cloning this repo as the project cwd also loads .opencode/skills and .opencode/agent/.

What is tailored per harness

	Claude Code	Grok Build	Codex CLI	Copilot CLI	OpenCode
Keep going	/loop	lead loop or gauntlet workflow	lead loop	lead loop	lead loop
Fan-out	ultracode + Agent	spawn_subagent (depth 1)	spawn_agent / wait_agent	task / /fleet	task
Resume builder	Agent agentId / SendMessage	resume_from	resume_agent	resume task or new task with gap	new task with gap
Fresh critic	new Agent	new spawn_subagent	new spawn_agent	new task	new task

Shared (all harnesses): named fetchable bar, blind binary critic, no round-count exit, gauntlet-progress.md.

/loop and ultracode appear only in the Claude skill and references/run-claude.md. Grok, Codex, Copilot, and OpenCode prompts tell the lead to keep looping with that harness's spawn tools instead.

Layout

skills/gauntlet-loop/           dispatcher skill (plugin install)
  references/loop.md            bar, flow, voice
  references/builder.md
  references/critic.md
  references/run-claude.md      /loop + ultracode + Agent
  references/run-grok.md        spawn_subagent + workflow
  references/run-codex.md       spawn_agent
  references/run-copilot.md     task + /fleet
  references/run-opencode.md    task
  workflows/gauntlet.rhai       Grok Build loop
harnesses/<name>/SKILL.md       pre-bound skill for clone-as-project
.claude/skills/                 Claude discovery
.agents/skills/                 Codex discovery
.github/skills/                 Copilot discovery
.opencode/skills/               OpenCode discovery
.grok/skills/                   Grok discovery
plugins/gauntlet-loop/          installable plugin
  agents/<name>.md              Claude + Grok agent definitions
  agents/<name>.agent.md        Copilot CLI agent definitions
  .codex/agents/<name>.toml     Codex agent definitions
scripts/check.py                cross-harness correctness checks

Both agent flavours share plugins/gauntlet-loop/agents/ because Copilot CLI discovers *.agent.md there. Claude Code globs *.md, which swallows *.agent.md too, so the two files collide on their name: field — and the Copilot flavour, whose tools: names Claude cannot resolve, wins and leaves the agent with no tools. .claude-plugin/plugin.json therefore lists the two .md files explicitly instead of letting Claude scan the directory. Keep that list in sync if you add an agent, and run python3 scripts/check.py.

How it works

1. You give a goal.
2. It offers 2 or 3 bars. Named, fetchable by this session, comparable.
3. You pick one. It writes ~150 words and stops.
4. You paste it, or say run it. Builder/critic pairs until the critic picks ours.

The critic is a separate agent with fresh context. It opens the actual output, puts it next to the bar with the labels stripped, and picks one. Not a score.

What breaks it

• A vague bar.
• The builder judging its own work.
• A soft critic (scores instead of a pick).
• A fixed round count.
• Telling a non-Claude session to /loop or ultracode.

Credit

Technique: Matt Shumer / Claude of Duty.

Reusable skill this adapts: Jay E / RoboNuggets.

License

CC BY 4.0. Attribution in NOTICE.md.