← Back to the collection

gauntlet

The Gauntlet Loop as a native skill. Build a piece, capture what it really produces, put it next to a reference that is already excellent, and have a fresh, harsh critic pick the better one without knowing which file is which. The single gap the critic names goes to the next builder. Repeat until the piece wins.

The method is three things: fresh context for every builder and critic, a harsh critic prompt, and two unlabeled files. The plugin bakes that in so nobody has to paste a prompt or remember the rules.

Install

claude plugin install gauntlet@jtsternberg

It pulls in hotline as a dependency.

Skills

Skill	Claude Code	Codex	What it does
`setup`	`/gauntlet:setup [goal]`	`$gauntlet:setup`	Interviews for the goal, the reference bar, the criteria and the capture harness; writes and validates `<workspace>/.gauntlet/<run-id>/gauntlet.json`
`run`	`/gauntlet:run <spec> [--status\\|--stop\\|--resume]`	`$gauntlet:run`	Drives the loop. User-invoked only, because a run spends sessions

How a run works

init cuts gauntlet/<run-id> from the base branch in its own worktree, copies the run dir there, and files one bd task in the project if it uses beads.
The lead splits the work into the smallest pieces a critic can judge alone.
Each wave builds the open pieces one at a time (each a fresh hotline callee in the worktree, committing with a Gauntlet: <piece> r<n> run=<id> trailer), capturing each right after its commit.
Critics judge in parallel. Each gets a fresh $TMPDIR dir holding only A.<ext> and B.<ext> (assignment by seeded coin flip, key kept outside that dir), is dialed from that dir, and returns a JSON verdict. The script maps it back and commits the run dir with the verdict in the trailer.
Stop rules fold the append-only ledger.jsonl: wins_to_close wins close a piece, plateau_rounds losses plateau it, plus per-piece and total round caps, a wall-clock cap, and a STOP file. After a wave that changed enough pieces, and once before the end, a smoothing pass makes the pieces cohere; closed pieces are re-judged and a loss reopens them.

progress.html in the run dir is rewritten after every event. Gauntlet never pushes.

Harnesses

A harness is how a piece becomes one file: capture.command plus a medium hint (image, text, or opaque, which forces rubric judging). Two reference harnesses ship in references/: a Playwright element screenshot (harness-image.md) and a markdown-section copy (harness-text.md). Field reference: references/spec.md.

judge.mode is the user’s choice: blind-pairwise (default) or rubric (labeled, scored 1–10, a win is 7+ on every criterion) for references too recognizable to blind. judge.double: true judges twice with A/B swapped.

Requirements

hotline with its herdr transport (declared as a plugin dependency), and herdr. The engine finds hotline via GAUNTLET_HOTLINE_ROOT, the installed plugin, a sibling plugin dir, or ~/Code/claude-plugins/plugins/hotline, and refuses to run without herdr support.
git and Node 20+. bd is optional.
Every builder and critic runs with HOTLINE_DANGEROUSLY_SKIP_PERMISSIONS=1, because an unattended pane must not stall on a permission prompt.

Design record: docs/plans/2026-10-26-gauntlet-plugin-design.md.