template-audit

Status: shipped-beta · Ships to: core tier arm (GETFF_SKILLS_CORE, setup.d/lib.sh:61) · Fires at: Claude Code skill auto-activation on the trigger phrases in its frontmatter description

What it is

A skill that audits the docs templates the framework renders (e.g. a project's AGENTS.md) via a local advisory review, in two steps. Its own self-description: a "Session-bound advisory audit. FREE under Claude Code subscription. No API key. Not blocking." (.claude/skills/template-audit/SKILL.md:13). It is advisory — it reports findings for the session to act on; the deterministic enforcement of the same templates stays in a CI gate the skill names but does not own.

How it works

Per its ## Procedure section (.claude/skills/template-audit/SKILL.md:15):

1. Step 1 — deterministic probes (CI-equivalent). Run npm --prefix packages/core run test:template-render; "If this fails — stop and fix before Step 2" (.claude/skills/template-audit/SKILL.md:20,23).
2. Step 2 — LLM advisory checks (session-bound, P2/P3/P5). "Ask the current Claude session (no API call)" (.claude/skills/template-audit/SKILL.md:25), starting with "P2 — Paraphrase fidelity: Does rendered AGENTS.md convey" the rule's meaning (.claude/skills/template-audit/SKILL.md:29).

A ## Promotion trigger section (.claude/skills/template-audit/SKILL.md:38) closes the skill. Harness posture is declared portable: "prose advisory audit checklist; no harness primitives" (.claude/skills/template-audit/SKILL.md:6).

Satellites & companions

• USES: the deterministic template-render audit in packages/core (npm … run test:template-render, Step 1) — the skill's header points there as the "Deterministic CI gate" while disclaiming ownership of it (.claude/skills/template-audit/SKILL.md:11).
• ADAPTS: the local advisory-review pattern (session-bound, no paid API) shared by the core-tier skill family.
• ADDS: the trigger vocabulary for template audits ("template, audit, render, generated docs, AGENTS.md, paraphrase, cue placement, …", SKILL.md:3) and the two-step ordering that gates the advisory pass on the deterministic one.
• Census family satellites: E (templates + vendored renders) — the surfaces this skill audits.

Anchors

• .claude/skills/template-audit/SKILL.md:2 — name: template-audit
• .claude/skills/template-audit/SKILL.md:3 — description: 'Use when auditing rendered templates via local advisory review. Triggers: template, audit, render, generated docs, AGENTS.md, paraphrase, cue placement, local advisory, template-render, audit-template.'
• .claude/skills/template-audit/SKILL.md:6 — <!-- @harness-posture: portable — prose advisory audit checklist; no harness primitives -->
• .claude/skills/template-audit/SKILL.md:13 — Session-bound advisory audit. **FREE under Claude Code subscription.** No API key. Not blocking.
• .claude/skills/template-audit/SKILL.md:20 — npm --prefix packages/core run test:template-render
• .claude/skills/template-audit/SKILL.md:25 — **Step 2 — LLM advisory checks (session-bound, P2/P3/P5)**
• setup.d/lib.sh:61 — GETFF_SKILLS_CORE="template-audit ai-doc rule-research rule-tests"

All anchors at framework pin b069c59328aa8e08671d3a19fd25efa618dfc1c5 (fetched 2026-09-11 from artyhoo/getff staging; ancestry of census pin a1337cb301 verified).