Forge Gauntlet

Open quality-loop runtime for Matt Shumer’s gauntlet method (guide · Claude of Duty): named fetchable bar → fan-out builders → separate harsh critic → blind A/B → loop until win or you stop.

Continuity: docs/HANDOFF.md · Technique: docs/TECHNIQUE.md · Site: gauntlet-runtime.vercel.app

Quick start

npm install
npx playwright install chromium   # optional real screenshots
npm run gauntlet -- propose "a dark athletic landing page for a running brand"
npm run gauntlet -- meta --goal "…"            # Shumer meta-prompt
npm run gauntlet -- compose --bar a --goal "…" --agent claude-code
npm run gauntlet -- run --bar a --goal "…"
npm run gauntlet -- status
npm run gauntlet -- stop
npm run demo
npm run web    # http://localhost:8787 — local run UI

# Marketing site
cd site && npm install && npm run dev

CLI

Command	Purpose
propose "<goal>"	Offer 2–3 named bars, then stop
meta --goal …	Shumer meta-prompt → draft a short aim prompt
compose --bar … --goal …	Orchestrator system prompt (paste into Claude/Codex)
compose --plan file.md|html	Extract criteria/gates; emit prompt + ≤3 gap bullets
run --bar … --goal …	Validate bar, ledger, builder↔critic loop
status / stop / resume	Live progress, human brake, continue
shot --url … / compare a b	Apex-style named captures + smoke compare
validate-bar / demo	Health check / end-to-end demo

Useful run flags: --mode apex, --climb, --vision-critic, --llm-critic, --dispatch-only, --max-usd, --max-tokens, --preview-url, --agent cursor|claude-code|codex, --implementer …

Ledger: runs/<id>/ → progress.md, workbench.md, ORCHESTRATOR.md, CONTRACT.md (apex), evidence/, dispatch/.

Skill

mkdir -p ~/.cursor/skills/gauntlet && cp -R skill/* ~/.cursor/skills/gauntlet/
# npx skills add <this-repo>  → skills/gauntlet/

Triggers: /gauntlet, “beat this bar”, “compose gauntlet”, “gauntlet this plan”.

Modes

• Standard — UI/sites/writing/code; exit when pieces win blind + smoothing, or you stop.
• Apex / climb — long runs (needs Claude Max / Codex plan). Writes CONTRACT.md. Human or budget is the real brake.
• Compose / meta — paste into another agent session; runtime keeps ledger + evidence when you run.

Hard rules

1. Critic never sees builder rationale; fresh critic every retry.
2. Verdict binary: ours or bar + one gap. No 1–10.
3. Bar evidence fetched. Labels stripped.
4. Exit = wins + smoothing, human stop, or budget. No default round cap.
5. Human gates and budgets outrank “keep going.”

Attribution

• Technique: Matt Shumer · somethingbig.ai/gauntlet-loop · Claude of Duty
• Role-split / gates: NicholasSpisak/gauntlet-loop
• Apex capture lessons: jolbol1/apex-gp

License

MIT