Agent skill · Claude Code + Codex + 70 more

Work doesn't pass until a blind critic says so.

Gauntlet Loop turns any plan, spec, or objective into a copy/paste system prompt where one agent orchestrates, one implements, and fresh-context critics audit the evidence — never the claims.

$npx skills add NicholasSpisak/gauntlet-loop

MIT · NicholasSpisak/gauntlet-loop · installs via the skills CLI

workbench.md

item	iter	verdict
auth-flow	×2	PASS
rate-limiter	×1	PASS
rollback-path	×4	PASS
idempotency	×3	FAILPASS
prod-promotion	—	AWAITING HUMAN

Every verdict comes from a fresh critic that never saw the builder's reasoning. Human gates outrank the loop.

§1

Three roles, one hard boundary

Single-agent "do it perfectly and check your work" prompts fail twice: the implementer grades its own homework, and iterative reviewers grade improvement instead of the bar. Gauntlet Loop separates the roles structurally, so there is nothing to negotiate.

Orchestrator / Critic

Claude

Decomposes the plan into the smallest separately-judgeable pieces, writes delegation contracts, adjudicates evidence, owns every gate.

Never implements.

Implementer

Codex

Produces every artifact through compact XML-block contracts: task, output shape, follow-through defaults, verification loop, action safety.

Never grades its own work.

Blind critics

Fresh-context sub-agents

Hostile auditors. They receive the artifacts and the verbatim acceptance criteria — never the implementer's summary, and never a draft they've seen before.

One unproven criterion = FAIL.

§2

The loop

ContractExtract one task's criteria, evidence requirements, and gate. Write the Codex delegation prompt.
ImplementCodex works until its own verification loop passes — read-back after every write, never success from a write response alone.
Blind auditA fresh critic gets artifacts + verbatim criteria, nothing else. It compares against the reference bar side by side and demands direct evidence per criterion.
IterateOnly the critic's findings go back to Codex. Every retry is judged by a new critic instance with clean context.
Adversarial second opinionSecurity-sensitive or irreversible items get an independent challenge pass. Every surviving finding is a FAIL.
Mark and logOnly after critics pass: mark complete, record evidence in the workbench ledger, move on. A final smoothing pass reviews the integrated whole.

No arbitrary iteration cap — the loop ends when the bar is met, a human gate is reached, or resources run out. Looping past a human gate is a critical failure, not persistence.

§3

What the generated prompt contains

Mission

The destination, never the implementation — architecture belongs to the implementer.

The bar

The document's own criteria plus concrete reference exemplars for blind side-by-side comparison. The work must win or tie.

Role split

The hard boundary: orchestrator delegates and judges, Codex builds, critics stay blind.

Delegation contract

XML blocks tuned for GPT-5.4-class implementers: task, verification_loop, action_safety, STOP conditions.

Critic contract

The hostile-auditor prompt. "Merely unimpressed" is a FAIL.

Fan-out map

Gated-sequential phases vs. surfaces safe to verify in parallel.

Hard stops

Human approvals and fail-closed safety rules that outrank the loop — sign-offs are never self-approved.

Definition of done

Every criterion critic-verified, smoothing pass complete, every pending human gate listed.

§4

Usage

# install (globally, to Claude Code and Codex)
$ npx skills add NicholasSpisak/gauntlet-loop -g -a claude-code -a codex

# then, in your agent — any of:
/gauntlet-loop build a CLI that syncs bookmarks across browsers
/gauntlet-loop docs/plans/payment-migration.md
/gauntlet-loop specs/redesign.html

Inline objectives get derived acceptance criteria marked DERIVED for your confirmation. Markdown and HTML plans are treated as the source of truth — their checkboxes, SLAs, and "never do" rules are extracted verbatim. The output is one fenced code block: the complete system prompt, ready to paste into a fresh session.

§5

Lineage

This skill extends Matt Shumer's Gauntlet Loop technique — give the agent a real bar it cannot talk its way around, split builder from critic, and loop with no arbitrary endpoint — and adds delegation contracts for Codex-class implementers, fresh-critic-per-retry adjudication, and human gates the loop can never override.