Gauntlet

Files (1)

• SKILL.md

SKILL.md (first lines)

Triggers the **gauntlet-lead** subagent in a forked context.

Multi-agent quality harness (Matt Shumer's gauntlet-loop technique): the LEAD sets a real
reference bar and splits the goal into independently gradeable units; BUILDERS produce artifacts
in clean context; BLIND CRITICS — a fresh context per unit per round, seeing only the artifact,
the bar, and the exemplar — grade with evidence; failures loop back until every unit passes, two
rounds stall, or the budget runs out.

**The two inviolables:** the agent that builds never grades its own work, and a critic that saw a
previous draft never grades the retry. Prior critiques go to the builder as fix input, never to
the next critic.

**The bar must be real:** a named product screen, a test suite + threshold, a model doc, a
measured baseline — written to `docs/gauntlet/BAR_<slug>.md` before round 1. "Make it amazing"
gets `BLOCKED: no real bar`. Aspirational bars are legitimate; uncomparable ones are not.

**Protocol:** `agents/shared/GAUNTLET_LOOP.md` — including boundaries vs `/challenge` (fact-checks
claims), Fix-Verify (closes known defects), and `/design-iterate` (token conformance against our
own spec — cheaper when the bar is internal).

**Outputs:**
- `docs/gauntlet/BAR_<slug>.md` — exemplar, per-criterion checks, budget
- `docs/gauntlet/GAUNTLET_<slug>.md` — round log, PASS evidence, which exit rule fired, below-bar residuals

From bpmforge/attest-claude, Apache-2.0.

More from bpmforge/attest-claude