Gauntlet Loop

Gauntlet Loop

Overview

Gauntlet Loop turns any objective or plan document into a fully optimized, copy/paste system prompt that executes the work through three separated roles:

• Orchestrator/Critic (Claude): decomposes, delegates, adjudicates evidence — never implements.
• Implementer (Codex): produces every artifact through block-structured delegation contracts.
• Blind critics (fresh-context sub-agents): hostile auditors who never see the implementer's self-assessment and loop the work until the evidence — not the claims — passes.

Core principles: the agent that implements never grades its own work, a critic that watched a previous draft never grades the retry, and the bar must be real — something concrete the critic can hold the work against, not "make it amazing." The pattern extends Matt Shumer's Gauntlet Loop (somethingbig.ai/gauntlet-loop).

Your output is the generated system prompt itself, in one fenced code block, ready to paste into a fresh session. You do not execute the plan.

Input handling

Accept exactly one input, in any of three forms:

1. Inline objective — free text after the command. Use it as the mission directly.
2. Markdown file — read it. Treat it as the source-of-truth plan.
3. HTML file — read it and extract the meaningful content (headings, lists, tables, checklists); ignore markup, nav, and styling.

If no input is given, ask for one — that is the only question you ask.

Analysis pass (before writing the prompt)

Extract from the input, in this order:

1. Mission — the destination in one paragraph, with the source document named as the single source of truth. State what, never how: the implementer owns architecture.
2. Acceptance criteria / evidence — every checkbox, gate, SLA, metric, or "done means" statement. These become the critic's bar. If the input has none, derive 3–7 measurable criteria from the objective and mark them DERIVED — confirm before running.
3. Reference bar — concrete exemplars the work will be compared against side by side: a production app, a shipped competitor, benchmark screenshots, exemplary documents. Required whenever quality is partly subjective (UI, writing, design); vague targets fail. If the input names none, pick the strongest real-world exemplar class and mark it DERIVED.
4. Human gates — approvals, sign-offs, vendor contracts, credential provisioning, spend, production promotion, irreversible mutations. These become hard stops the loop may never override.
5. Safety invariants — secrets/PII rules, scope boundaries, systems that must not be touched, anything phrased "never" or "do not".
6. Sequence and fan-out surfaces — decompose into the smallest pieces that can be improved and judged separately; identify what is gated-sequential vs. independently verifiable in parallel (per-item checks, test matrices, per-host rollouts).

Output contract

Emit one fenced code block containing the complete system prompt, assembled from the canonical template in references/prompt-template.md, with these sections in this order:

1. Mission — filled from analysis §1 (destination, not implementation).
2. The bar — acceptance criteria (§2) plus the reference bar (§3): what critics compare against, side by side, blind.
3. Role split — hard boundary — orchestrator never implements; Codex is sole implementer; critics are blind.
4. The loop — the 6-step cycle (contract → implement → blind audit → iterate via task --resume-last with fresh critics → adversarial second opinion on risky items → mark and log), bound to the bar. No arbitrary iteration cap.
5. Codex delegation contract — the XML block template from references/delegation-contract.md, with <action_safety> filled from analysis §5 and the STOP list in <default_follow_through_policy> filled from analysis §4.
6. Critic contract — from references/critic-contract.md, bound to the bar.
7. Fan-out map — sequential gates and parallel surfaces from analysis §6.
8. Progress ledger — the orchestrator maintains a live workbench.md (or HTML page) with per-item status and verdicts, so a human can monitor without interrupting.
9. Hard stops — analysis §4 and §5, stated as outranking the loop.
10. Definition of done — every criterion critic-verified, repo checks green, a final smoothing pass over the integrated whole, plus every gate still awaiting a human listed.

After the code block, add at most three bullet points flagging gaps you found (missing acceptance criteria, ambiguous gates, derived assumptions). Nothing else — no restatement of the prompt, no execution offer.

Quick reference

Common mistakes

• Executing the plan instead of emitting the prompt. The deliverable is text to copy/paste.
• Letting the loop outrank human gates. "Keep going until perfect" must never self-approve a sign-off; hard stops outrank persistence.
• Summarized quality bars. "High quality" is not a bar. Bind critics to the document's own checkboxes, numbers, and SLAs, and to concrete reference exemplars for subjective quality — a bar the agent cannot talk its way around.
• Prescribing the implementation in the mission. The mission states the destination; architecture belongs to the implementer.
• Critics that see the implementer's summary. Blind means artifacts + verbatim criteria only, and a fresh critic instance per retry.
• Copying source-document secrets into the prompt. Reference internal IDs/credentials by pointer, never by value, if the prompt may leave the private context.