agentleFS
unlockers-aiunlockers-ai/unlockers-skills-library/skills/review-ai-writing/SKILL.md

review-ai-writing

Skill4 starsChanged 3 months agoMITView on GitHub

Review prose (blog post, article, doc, README, LinkedIn post) for signs of AI writing — the tells catalogued in Wikipedia's "Signs of AI writing", plus French-specific ones — and propose rewrites in the author's voice as numbered diffs the author approves or refuses one by one. Use when asked to review a text for AI-isms, "AI-ismes", "slop", "does this sound like ChatGPT", "make it sound less AI", or before publishing a text drafted with an LLM.

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Leave alone what the author wrote by hand: If they rewrote passages in this session or say a passage is theirs, flag only clear grammar …

    line 23
  2. Distinctive:

    Apply: only the numbers the author accepts.

    line 32
  3. Distinctive:

    Write the rewrites in the text's language, in the author's register: shorter, more concrete, one example rather than an abstract category.

    line 95

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Review AI writing

Find the passages a careful reader would notice as machine-written, and propose a rewrite for each that keeps the meaning and sounds like the author. The author decides; this skill does not edit the text until told which suggestions to apply.

Reference

Wikipedia's Signs of AI writing is the catalogue. The tells below are the ones that apply to prose in general; the page has more (markup, citations) that rarely matter outside Wikipedia.

Process

  1. Read the whole text, and something the author wrote before (an earlier post, their emails) for voice. An earlier text may itself be AI-assisted: use it as a rough reference, not ground truth.
  2. Leave alone what the author wrote by hand. If they rewrote passages in this session or say a passage is theirs, flag only clear grammar errors there.
  3. Check every claim tied to a source. Open the link and read what it actually says. A figure that is real but comes from a study about something else, an "explanation la plus répandue" with no source, a study that measures X presented as proving Y: these are the most damaging tells, because a reader who clicks loses trust in the whole text.
  4. Apply the tells below. Be selective: report what a careful native reader would notice, not every stylistic choice. One "crucial" is coincidence; a cluster is a tell.
  5. Report as numbered diffs in reading order (see Report).
  6. Apply only the numbers the author accepts. When they redirect one ("for 4, rather say…"), write the new version from their words and show it before committing.

The tells

Content

  • Inflated significance — tying a detail to a broader theme: "marque un tournant", "reflète une tendance de fond", "pivotal moment", "reflects broader…".
  • Trailing -ing / participle analysis — "…, soulignant l'importance de…", "highlighting…, underscoring…, fostering…".
  • Vague attribution — "les experts", "l'explication la plus répandue", "observers have noted", with no one named and no link.
  • Over-claiming — a statement more certain than its source; a hedge immediately retracted ("je le prends comme un indice. Il va tout de même dans le sens de…").
  • Formulaic conclusion — "despite challenges… the future looks…", or a closing paragraph that re-lists every section.
  • Promotional tone — "incontournable", "riche", "au cœur de", "groundbreaking", "vibrant", "boasts".

Language

  • AI vocabulary in clusters — EN: delve, crucial, pivotal, tapestry, underscore, showcase, intricate, landscape, leverage. FR: crucial, essentiel, enjeu(x), levier, dispositif, trajectoire, écosystème, naviguer, plonger dans, "un véritable…".
  • Avoiding plain verbs — "constitue", "s'inscrit dans", "serves as", "stands as", "features" where "est" or "a" would do.
  • Negative parallelism — "ce n'est pas X, c'est Y", "il ne s'agit pas de… mais de…", "not just X, but Y", "no X, no Y, just Z".
  • "Y plutôt que X" used to manufacture insight from a reversed claim.
  • Rule of three — three adjectives, three short phrases, three examples, by reflex.
  • Balanced pairs — "c'est un avantage… mais cela apporte aussi…", "loin d'être parfait, mais il constitue…": every point weighed on both sides.
  • Calques from English — "adresser un problème", "impacter", "inconsistance" (for inconsistency), "construire l'intuition de quand", "faire du sens".

Structure and rhythm

  • Signposting — "Ici, je regarde…", "Voyons maintenant…", "Il existe une technique pour…", "Let's dive in".
  • Mechanical transitions — a paragraph opening with a sentence that only announces the topic ("Pour la délégation, c'est un obstacle de plus.").
  • Tidy closers — a final sentence restating the paragraph, or tying it back to an earlier section ("…alimente directement la peur du premier point.").
  • Uniform paragraphs — every paragraph the same length and shape.
  • Generic headings — "Ce qui se joue", "Conclusion", "Key takeaways", title case, "X and Y" pairs.
  • Formatting leftovers — bold key terms in running text, "Label: description" bullet lists, emoji before headings, a horizontal rule between every section, chatbot phrases ("J'espère que cela vous aide", "Certainly!"), placeholders ("[Votre nom]").

Report

A ranked summary first: a one-line verdict (low / medium / high AI flavour, and where the tells cluster), then the findings in reading order, numbered so the author can answer "ok 1-5, 7; refuse 6". Each finding:

N. Section: the tell, in a few words (line)

- the exact current sentence(s), only those that change
+ the proposed rewrite

A finding that touches separate places (a heading and a paragraph) gets one diff block per place under the same number, and says the parts can be taken separately.

Write the rewrites in the text's language, in the author's register: shorter, more concrete, one example rather than an abstract category. Keep links, figures and formatting intact unless the finding is about them.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.