Reconcile agent-audit skill with create-skill conventions #94

Context

skills/agent-audit/ was authored in one go without first loading the canonical skill-authoring skills:

• skills/create-skill/SKILL.md (this repo): frontmatter rules, directory layout, AGENTS.md integration.
• obra/superpowers third-party skills: write-a-skill, writing-skills — progressive disclosure, bundled-resource patterns.

The skill currently works end-to-end (verified against Claude Code 2.1.160, OpenCode 1.15.13, Codex 0.136.0, pi 0.78.0; Cursor reported as not installed), but its structure may diverge from the house pattern those skills prescribe.

Files to audit

• skills/agent-audit/SKILL.md — frontmatter, section structure, progressive disclosure
• skills/agent-audit/audit.sh — bundled-script conventions
• skills/agent-audit/lib/claude-loopback.mjs
• skills/agent-audit/lib/opencode-plugin.ts
• skills/agent-audit/lib/pi-dump.mjs

What to do

1. Load create-skill, write-a-skill, and writing-skills and read each.
2. Diff their prescriptions against the current skill: frontmatter fields, section headings/order, bundled-resource layout (lib/ vs other convention), progressive disclosure depth, self-improvement loop wording.
3. Update skills/agent-audit/* to match. Keep behaviour intact: a re-run of ./audit.sh must produce the same five-row table.
4. If any convention is ambiguous or contradicts another, surface it for discussion before changing.

Out of scope

• Adding a new client. Cursor's "not installed" path stays as-is.
• Swapping the chars/4 token estimate for a real tokenizer (already noted in the skill's self-improvement loop).