skill-audit

Audit a single skill as a closed world — its SKILL.md plus everything under references/ (and bundled assets), analyzed together — and report where it contradicts itself, repeats itself, or carries detail that doesn't earn its tokens. This is internal-consistency review, not correctness review: it never judges whether the skill's advice is right about any codebase, only whether the skill is coherent and lean against its own stated purpose.

The unifying question for every unit of content is "does this earn its tokens?" — and the standing danger is becoming a terseness-maximizer that strips the hard-won nuance preventing known failures. A dedicated adversarial guard exists to stop exactly that (see references/load-bearing-guard.md).

What it does (and what it does not)

It does: decompose the target into a typed unit ledger, run three lens reviews (contradiction / redundancy / dilution) and an adversarial load-bearing guard over that ledger, merge findings by unit ID, and emit a prioritized, apply-able report.

It does not: compare against upstream (that is skill-diff's job), audit more than one skill at a time, review arbitrary markdown, or auto-apply contradiction fixes. Contradiction detection is high-confidence closed-world; dispositions and load-bearing verdicts are inferred judgments — the report states this boundary.

How this works

The orchestrator (this file) owns the whole pipeline in one context. The lenses and guard are prompt files under references/ that the orchestrator reads and runs as sequential in-context passes over a single shared ledger — they are not dispatched to subagents. Because every pass scores the same unit IDs by construction, merge-by-ID has no cross-context drift. This is the brave-breakdown read-references-inline convention, not a subagent fan-out.

Phases, in order: