Ai written check
Keep AI-assisted writing honest and yours: an AI-written check, a cringe check, and a full-review orchestrator (anti-inflation + fact passes). Composable Claude Agent Skills.
npx -y skills add cabbagecachekid/red-pen --skill ai-written-checkAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Detects the mechanical tells that make professional prose read as AI-generated — em dashes, triple-list anaphora, "not X but Y" antithetical pairs, spatial and economic metaphors for abstract ideas, "architect" as a verb, engineered sentence cadence, mid-paragraph register swerves, builder/startup slang, and unfalsifiable universal claims ("everyone", "across the board"). Use when reviewing or editing a resume, cover letter, portfolio page, case study, proposal, bio, or any human-facing copy, or when the user asks for an "AI-written check", "AI tells", says copy "sounds like AI", or asks to make writing sound less generated.
SKILL.md
4.8 KB, as published. Nobody here has run it
AI-Written Check
Scan human-facing prose for the mechanical patterns that signal machine authorship, and propose a concrete rewrite for each. This is a prose check — it looks at sentence mechanics and word choice, not at tone or positioning (that is the cringe-check skill).
What this is for. This is not about hiding that AI was involved. It assumes the healthy case: you're collaborating with Claude to refine, expand, or draft your own writing. The job is to keep the result aligned with your intent and your voice — so the words still sound like you wrote them, not like a model smoothed them into generic copy. None of these skills are for making AI do the writing for you.
How to run it
- Read the copy once for sense.
- Walk the checklist below. For every hit, output: the rule name, the offending text quoted, and a specific rewrite — not just "this sounds AI."
- Respect the thresholds. Some patterns are fine once and only become a tell when repeated. Count occurrences before flagging.
- End with a one-line verdict: clean, light touch-up, or heavy.
Never flag without a rewrite. The point is to fix the prose, not to label it.
The checklist
| Tell | Threshold | Fix |
|---|---|---|
| Em dash (—) | The #1 tell. Flag almost every one in body copy. | Period, colon (sparingly), or a shorter sentence. Acceptable only as a structural separator in a header. |
| Triple-list anaphora ("about money, about trust, about control") | 1 per page is fine; 2+ is a tell. | Cut to a single clause or break into separate sentences. |
| "Not X, but Y" / "X, not Y" antithetical pairs | 1 per piece is fine; more is a tell. | State the point directly without the contrast scaffold. |
| Spatial metaphors for abstract concepts ("corners of their financial lives", "the territory of taxes", "the current frontier") | Flag each. | Use the concrete noun the metaphor is standing in for. |
| Economic/cost metaphors for non-economic things ("make rigor the cheap option", "the time tax", "the heavy lift") | Flag each. | Say the literal thing. |
| "Architect" as a verb | Flag each. | "Design" or "build." |
| "Across" overuse ("across the funnel", "across products", "across teams") | Flag clusters. | Name the specific thing, or drop the word. |
| "From X to Y to Z" career/idea arcs | Flag when a plainer sentence works. | Rewrite as a direct sentence. |
| Engineered sentence cadence (deliberate alternation of very short and very long sentences) | Flag when the rhythm feels manufactured. | Let sentence length follow meaning, not a beat. |
| Voice-register swerve | Flag any paragraph that opens concrete-and-human and pivots to abstract industry-speak. | Stay in one register; the swerve is the tell. |
| Builder/startup subculture slang ("vibe coding", "shipping", "hot takes", "cracked at") | Flag each in formal copy. | A neutral substance-equivalent term. |
| Unfalsifiable universal claims ("everyone", "all", "every", "universally", "across the board") | Question each. | Scope it to the specific people/context actually observed. "Affects everyone" is itself an AI-pitch tell. |
| Generic placeholder / template copy (fill-in-the-blank headers, "Welcome to my portfolio", "I'm passionate about X", boilerplate section intros that could belong to anyone) | Flag each. | Replace with something specific and true to the author — their actual words, project, or point of view. Filler that "isn't the vibe" is a tell that the line was generated, not authored. |
Why these matter
These patterns are how a reader's pattern-matcher flips to "a model wrote this." Most are individually harmless — the damage is density. One em dash is invisible; six is a signature. Apply the thresholds rather than nuking every instance, or the rewrite reads as over-corrected.
For the full catalog with worked before/after examples and edge cases, see references/TELLS.md.
What this skill does NOT do
- It does not judge tone, arrogance, or positioning — use
cringe-check. - It does not verify facts or claims — that belongs to a full review (
full-review). - It does not rewrite wholesale. It flags and proposes; the author decides.