content-audit

A Claude skill that acts as a strict pre-release content auditor. Point it at a project folder and it reads the text your real users will see — landing pages, marketing sites, and web/desktop/mobile app UI — then reports everything that would make the product look unfinished, leaky, or obviously AI-written, with a concrete fix for each line.

It is deliberately not a grammar checker. Grammatically perfect copy can still be generic AI sludge, and clean-looking text can still hide a leftover TODO. The skill judges every line against the bar set by Apple, Stripe, Linear, and Notion.

What it catches

1. AI-like writing — buzzword benefit-claims and filler ("seamless experience", "powerful solution", "streamline your workflow", "can be used to"), template intros, repetitive structure, and low-information copy — in any language the product ships, not just English.
2. Developer-facing content leaked to users — TODO / FIXME / WIP, debug messages surfaced as UI, prompt/AI leakage, internal notes.
3. Internal implementation details — table/column names, stack traces, error codes, internal hostnames/URLs exposed in user copy.
4. Placeholder / unfinished content — lorem ipsum, dummy names/emails, "Coming soon" left indefinitely, untranslated strings, rendered null/undefined.
5. Low-quality UX copy — vague, ambiguous, overly technical, or value-free copy; error messages with no cause or next step; buttons that don't say what they do; weak confirmation / empty / loading states; tone that doesn't match the moment.

It audits the words, not the facts — whether copy reads naturally and well, not whether a price, model ID, route, or legal detail is correct. It won't assert that an external fact is wrong (it can't verify that from the copy); anything that looks factually off is flagged at most as "verify," never as a defect.

How it works

The flow is review-driven — the skill produces a report you review (and edit), and only applies fixes after you confirm:

1. Discover   → map what users see + detect which languages the product ships
2. Plan       → fan out per-language subagents when there are many locales (i18n)
3. Audit      → find issues across the dimensions above
4. Report     → write a reviewable Markdown report (path:line + revision + scores)
5. Apply      → ONLY after you confirm the (possibly edited) report

Two details worth knowing:

• Multilingual by design. Detection is language-agnostic. The skill finds every locale the product ships and audits each one against the same language-independent AI-writing patterns — empty benefit-claims, template intros, hype without proof, redundant politeness — judging each locale by its own native conventions rather than a translated word list; in testing it flagged generic German marketing copy with no German-specific rules. For products with many locales it audits them in parallel (one subagent per language), then merges the findings into one report.
• Scoped to user-facing text. Code, comments, commit messages, developer READMEs, and console.log/debugPrint lines are excluded from style review — unless that developer text has leaked into a surface a user can actually see, which is itself a finding.

The report

Findings are grouped (AI-like / Developer-Facing & Unfinished / UX Copy), each with the verbatim original, a path:line location, and a drop-in suggested revision. The report closes with an Overall Assessment including two scores:

• AI-Likeness Score (0–100) — higher means more AI-like (worse).
• Release Readiness Score (0–100) — higher means more ready to ship. A single critical leak (e.g. a shipped TODO or an exposed internal error) caps this at 60.

Install

As a packaged skill: install content-audit.skill through your Claude skills UI.

Manually: copy the content-audit/ folder (SKILL.md + references/) into your skills directory, e.g.:

~/.claude/skills/content-audit/

Usage

Just describe the task in natural language — the skill triggers on intent, not a specific command:

• "Audit the landing page in ./site before we launch — make sure nothing sounds AI-generated and there's no leftover developer text."
• "Pre-release content review over ./app. It's localized, so check every language we ship, and flag anything untranslated or placeholder."
• "Does this copy sound robotic? Find any placeholder text still showing in the UI."

The skill writes content-audit-report.md to your project root. Review it, tweak any revision or delete false positives, then tell Claude to apply the report and it will edit the referenced files (preserving interpolation placeholders and i18n keys).

Repository layout

content-audit/
├── SKILL.md                      # the skill: workflow + triggering
├── references/
│   ├── ai-tells.md           # AI-writing tells (language-independent)
│   ├── detection-rules.md    # user-facing vs code; dev-leak / placeholder / internal patterns
│   ├── ux-microcopy.md       # UX copy quality: errors, buttons, dialogs, empty states, forms
│   └── report-format.md      # output template + scoring rubric
├── LICENSE
└── README.md

License

MIT — see LICENSE.