audit-and-fix · git:20260722.45cbafa · 2026-07-22 · sha256 e156051f4232f936
audit-and-fix git:20260722.45cbafaA
Immutable. This exact content never changes and is served at /api/v1/blob/e156051f4232f936.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_bxhcnelxczxp32oq/e156051f4232f9366015c5c804d0338f76275727236ea83c46d86b98d6358540.json
Then /plugin install audit-and-fix@markdownregistry-bxhcnelxczxp32oq in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
---
name: audit-and-fix
description: Audit ALREADY-published content and fix low-quality items, for an agent built with content-agent-kit. Use when the user says "audit the published content", "review old posts and fix them", or "clean up the backlog". Rule-audit first, then subagent grading, then STOP for confirmation before mutating live content.
---
# Audit & fix (published content)
This is separate from daily creation and it mutates LIVE content — so it stops for human
confirmation before changing anything.
1. **Rule-based audit first (cheap).** Run the project's rule audit (e.g.
`node scripts/audit-quality.mjs`) → `audit_report.md`. Hard rules only: length, missing
fields, broken id references, duplicate/missing images. No AI needed.
2. **List** recent published items: `node scripts/list-published.mjs --limit 40` →
`audit_input.json`.
3. **Subagent audit (fan-out)** — one reviewer per item grades cover/images/content/logic
→ `{ id, ok, issues[] }`.
4. **STOP and present.** Write `audit_report.md` (failing items + proposed fixes) and show
the user. Do NOT change anything until they choose which items to fix.
5. **Fix confirmed items** — regenerate the bad part / new images, keep the `id` stable,
apply with `node scripts/update.mjs <fix>.json`. Re-audit to confirm clean.
6. **Report** → `brain/<id>/audit_fix_report.md`: fixed / skipped / dropped.
**Priority**: items visible to logged-out / free users first, then paid, then the long
tail. Work in batches of 20–30 (not >100 per run).