setup
/setup
Set up a Gauntlet Loop for a project: define the goal, the reference bar the work must beat, judging criteria beyond looks, how each piece is captured and judged (the harness), and the stop rules, then write and validate a gauntlet.json spec. Use for 'gauntlet this', 'set up a gauntlet', 'make this beat <reference>', 'push this past good enough', 'blind critic loop', 'keep iterating until it beats X'.
- Author
- jtsternberg ↗
- Repository
- claude-plugins ↗
- Licence
- MIT
- Last repository change
- 13.11.2026
Assessment and security
- Files in this package
- 2
- Scripts
- 0
- HTML comments
- 0
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 13.11.2026, without references to retired models or APIs.
MIT, open: use and modify while retaining author attribution.
What it does and when to use it
Set up a Gauntlet Loop for a project: define the goal, the reference bar the work must beat, judging criteria beyond looks, how each piece is captured and judged (the harness), and the stop rules, then write and validate a gauntlet.json spec. Use for 'gauntlet this', 'set up a gauntlet', 'make this beat <reference>', 'push this past good enough', 'blind critic loop', 'keep iterating until it beats X'.
Triggers from the description
Arguments
/setup [goal | path/to/gauntlet.json to revise]
Original author description
Set up a Gauntlet Loop for a project: define the goal, the reference bar the work must beat, judging criteria beyond looks, how each piece is captured and judged (the harness), and the stop rules, then write and validate a gauntlet.json spec. Use for 'gauntlet this', 'set up a gauntlet', 'make this beat <reference>', 'push this past good enough', 'blind critic loop', 'keep iterating until it beats X'.
How it works
- Input
- Where things go
- Interview
- Write and check
- Hand off
SKILL.md preview
---
name: setup
description: "Set up a Gauntlet Loop for a project: define the goal, the reference bar the work must beat, judging criteria beyond looks, how each piece is captured and judged (the harness), and the stop rules, then write and validate a gauntlet.json spec. Use for 'gauntlet this', 'set up a gauntlet', 'make this beat <reference>', 'push this past good enough', 'blind critic loop', 'keep iterating until it beats X'."
argument-hint: "[goal | path/to/gauntlet.json to revise]"
allowed-tools: "Read Write AskUserQuestion WebFetch Bash(node *) Bash(git *) Bash(gh repo view *) Bash(mkdir *) Bash(cp *)"
---
# Gauntlet: Setup
The Gauntlet Loop: build a piece, capture what it really produces, hand that and a
reference that is already excellent to a **fresh, harsh critic** that does not know
which is which, take the one gap it names, rebuild, and repeat until the piece wins.
Fresh context, a harsh prompt, two unlabeled files. That is the whole method.
This skill writes the spec the loop runs from. It builds nothing and judges nothing.
The run skill does that, and only when the user asks for it.
## Input
The request is `$ARGUMENTS`: either a goal in plain words or the path to an existing
`gauntlet.json` to revise.
Codex: if the token above is not substituted, read the goal or path from the user's message.
## Where things go
Everything lives in `<workspace>/.gauntlet/<run-id>/`: the spec as `gauntlet.json`,
references under `refs/`, and the harness's capture script if it has one. `<workspace>`
is the repo the builders will change (`git rev-parse --show-toplevel` from the current
directory, unless the user names another). Pick a short lowercase `run_id`, such as
`landing-2026-10-26`.
**Read `${CLAUDE_PLUGIN_ROOT}/references/spec.md`** for every field and its default.
Codex: substitute the installed gauntlet plugin directory for the path above.
## Interview
One question at a time, each with a recommended answer. (Codex: ask as plain chat
turns.) Skip any the request already answered.
1. **Goal.** What is the whole artifact for, and for whom? One or two sentences.
2. **The bar.** Which references must it beat? Files the user hands over, URLs to
capture, or a named bar ("Linear's homepage") you capture. Never generate a
reference yourself. A reference that is merely okay makes the loop worthless.
3. **Criteria.** Two to five questions the critic answers, each starting "Which…".
At least one must be about the artifact doing its job (clarity, conversion,
correctness), not only how it looks.
4. **Harness.** How a piece becomes one file the critic can read:
- a web page or component → the image reference harness. **Read `${CLAUDE_PLUGIN_ROOT}/references/harness-image.md`**.
- prose → the text reference harness. **Read `${CLAUDE_PLUGIN_ROOT}/references/harness-text.md`**.
- anything else → write a `capture.command` that writes one file to `{out}`; if no
blinding can hide which file is ours, set `medium: "opaque"` (forces rubric).
Codex: substitute the installed gauntlet plugin directory for both paths above.
Copy the harness script into the run dir when it uses one, then dry-run the capture
once on the current artifact (step below). Capture each URL reference with the same
harness into `refs/`, so ref and candidate come out of one pipeline.
5. **Judging.** `blind-pairwise` (default) or `rubric`. Recommend rubric only when a
reference is so recognizable that the critic would know which file is which.
`judge.double: true` judges twice with A/B swapped; it doubles critic calls.
A custom `judge.prompt` template must name the files and carry the answer shape:
`{{FILE_A}}` and `{{FILE_B}}` (pairwise) or `{{REFERENCE}}` and `{{CANDIDATE}}`
(rubric), `{{CRITERIA}}`, and `{{SCHEMA}}`. `{{STANCE}}` and `{{GOAL}}` are
optional. Copy `${CLAUDE_PLUGIN_ROOT}/references/judge-pairwise.md` (or
`judge-rubric.md` beside it) into the run dir and edit that.
Codex: substitute the installed gauntlet plugin directory for the path above.
6. **Git.** Check whether the remote is public:
```bash
gh repo view --json visibility -q .visibility
```
Captures are committed to the run branch by default. If the repo is public, ask
whether they may be (`git.commit_captures`); screenshots can carry data.
7. **Budget and stop rules.** Offer the defaults from `spec.md` in one line and ask
for changes. They can be raised mid-run by editing the run's `gauntlet.json`.
## Write and check
Write `<workspace>/.gauntlet/<run-id>/gauntlet.json`, then dry-run the capture and
validate. Use a piece id and capture arguments the split is likely to use.
Codex: this path resolves under Claude Code; substitute the installed gauntlet plugin directory.
```bash
PLUGIN_ROOT="${CLAUDE_PLUGIN_ROOT}"
node "$PLUGIN_ROOT/scripts/gauntlet.mjs" try-capture <workspace>/.gauntlet/<run-id>/gauntlet.json --piece <id> --args '{"selector":"#hero"}'
```
Read the file it reports. A capture that does not show the piece is a harness bug;
fix it before going on.
Codex: this path resolves under Claude Code; substitute the installed gauntlet plugin directory.
```bash
PLUGIN_ROOT="${CLAUDE_PLUGIN_ROOT}"
node "$PLUGIN_ROOT/scripts/gauntlet.mjs" validate <workspace>/.gauntlet/<run-id>/gauntlet.json
```
Fix every named field until it reports `"ok": true`.
## Hand off
Finish with the spec path and the command that starts the loop:
`/gauntlet:run <workspace>/.gauntlet/<run-id>/gauntlet.json` (Codex:
`$gauntlet:run <path>`). Say what a run spends: one builder session per piece per
round, one critic session per judgment (two with `double`), up to the budget.
Files in this package (2)
- SKILL.md 6 KB
- agents/openai.yaml 1 KB
Versions
| Repository | Score | Licence |
|---|---|---|
| This page · jtsternberg/claude-plugins | 100 | MIT |
| HwangSlater/harness-fit | 100 | MIT |
| nesquikm/dev-process-toolkit | 100 | MIT |
| userKael/rastro | 100 | MIT |
| als-apg/osprey | 100 | BSD-3-Clause |
| dedujeong/gemba | 100 | Apache-2.0 |
| eigenwise/eigenwise-toolshed | 100 | MIT |
| awrshift/agent-memory-kit | 100 | MIT |
| lifecycle-innovations-limited/claude-ops | 100 | MIT |
Database snapshot: 13.11.2026 · package version a1ccb2c3a439 (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 95gauntlet-loop
Turns any goal into one short, paste-ready gauntlet-loop prompt for OpenCode, then can run it with the task tool. Critics are new tasks every time. No /loop, no ultracode — those are Claude Code. Triggers on "/gauntlet-loop", "gauntlet loop", "gauntlet this". - 100gauntlet-loop
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.… - 100run
Run a Gauntlet Loop from a gauntlet.json spec: split the work into pieces, build each with a fresh hotline callee, capture the real output, judge it blind against the reference with an independent critic callee, commit each round, and loop until it wins, a stop rule fires, or… - 95gauntlet-loop
Turns any goal into one short, paste-ready gauntlet-loop prompt for Grok Build, then can run it with a lead-managed spawn_subagent loop. Depth-1 spawns. No /loop, no ultracode — those are Claude Code. Triggers on "/gauntlet-loop", "gauntlet loop", "gauntlet this". - 100gauntlet-loop
Beat a real-world quality bar with a Gauntlet Loop: take a GOAL and a REAL-WORLD EQUIVALENT with an inspectable reference pack, decompose into independent parts, assign specialist builders, and gate each part behind a separate fresh-context critic. Pass only if better than the… - 90gauntlet-performance
Use when a human asks how gauntlet runs perform across the recorded telemetry - explicit invocation only (/skill:gauntlet-performance [--dir <path>]... [--since <version>]).