Leave empty and the prompt makes the agent find its own concrete bar first.
I want you to [describe what you want to build]. Before building anything, find the bar: a concrete reference an agent can actually inspect and compare its work against — a real competitor product, reference screenshots, an exemplar document, a test suite with targets. Explain in one sentence why it is a useful bar, then judge every round against it. Vague standards like "make it amazing" don't count. Break the work into the smallest pieces that can be improved and judged independently — you decide the split, not me. For each important piece, fan out a builder subagent and a separate, genuinely harsh critic subagent with fresh context. The critic must never see the builder's reasoning or summaries — only the goal, the bar, and the real artifact. Each round, the critic inspects the real output — rendered screenshots, running code, actual test results, the finished text — and compares it side by side with the bar in a blind A/B. It states which one wins and names the single biggest remaining gap. If ours loses, the builder fixes that one gap and the piece goes through the gauntlet again. Hard cap: 3 rounds per piece. When a piece hits the cap, stop working on it and log the remaining gap honestly instead of burning more tokens. Report total rounds used per piece at the end. Maintain a live workbench.md progress page. After every round, append one line: piece, critic verdict, gap named, action taken. I will watch progress there — don't stop to ask questions unless you are truly blocked. When all pieces pass or cap out, run one final fresh agent over the assembled whole to smooth out inconsistencies between independently improved pieces before you finish.
Kit adds a bar cheat sheet, brake checklist, and a standalone critic prompt.
The agent gets a concrete reference it can inspect — real screenshots, a competitor site, a test suite. "Make it amazing" gives a critic nothing to compare against; a real artifact can't be talked around.
Separate agents. The builder makes the thing; a fresh-context critic blind-compares it against the bar and names the single biggest gap. The builder never grades its own work — it has seen every decision it made and will defend all of them.
Lose the A/B → fix that one gap → judge again. One gap per round keeps every iteration attackable. The method that built the viral "Claude of Duty" browser FPS ran this loop for hours across dozens of subagents.
The original method says "you are the brake" — and users have burned hundreds of dollars finding out what that means. Our template hard-caps rounds per piece and makes the agent report spend, so the loop ends on your terms, not your bill's.
The Gauntlet Loop is an agent-prompting pattern that swaps vague quality instructions for a concrete, inspectable bar. Instead of telling an agent to "make it amazing," you give it a real reference — actual Call of Duty screenshots in the original demo — and split the work between builder agents and independent critic agents that blind-compare output against that reference. The pattern went viral in July 2026 when a single 152-word prompt produced a 55,000-line browser FPS over an unattended multi-hour Claude Code run. It generalizes well beyond games: landing pages judged against best-in-category sites, articles judged against the top-ranking result, backend code judged against a test suite.
The generator assembles your prompt from six template blocks, entirely client-side. Block one states your goal verbatim. Block two sets the bar: your reference if you provided one, otherwise an instruction that makes the agent find and justify a concrete bar before building anything. Block three tells the agent to split work into the smallest independently-judgeable pieces and assign each a builder plus a fresh-context critic — the Codex variant first writes the full spec to goal.md as a durable source of truth. Block four encodes the blind A/B rule: critics inspect real artifacts (screenshots, running code, test output), never builder summaries, and name exactly one biggest gap per round. Block five is the brake — a hard per-piece round cap with honest gap reporting when the cap is hit. Block six adds the optional workbench.md live progress page and a final smoothing pass that reconciles independently-improved pieces.
A developer wants a polished browser game prototype and generates a prompt with Mario Kart screenshots as the bar, a 5-round cap, and the smoothing pass enabled.
A founder building a landing page sets the bar to the best competitor site in the category and lets a critic blind-compare Playwright screenshots each round.
A technical writer targets the current top-ranking article for a query as the bar, so the critic judges which text better answers the searcher.
An engineer caps rounds at 2 for a refactoring task, using the loop purely as a disciplined second-pass polish rather than an open-ended run.
Describe what you want the agent to build in plain language
Add a concrete quality bar — a real product, screenshots, or test suite the critic can compare against (or let the agent find one)
Pick your harness (Claude Code or Codex) and set a round cap per piece
Copy the generated prompt, paste it into your agent, and watch the workbench.md progress page
We packaged the .claude config that runs this site — 26 specialist agents, 14 workflow skills, and 6 rule files from 31 real production incidents. From $9.
Get the Claude Code Production Pack — $29