the skill
Avoid AI Tells
Write prose that does not pattern-match to AI-generated text. Apply this skill whenever the deliverable is written English for a reader — articles, blog posts, essays, client reports, decks, emails, Slack messages, documentation, marketing copy, bios, summaries, release notes, commit messages, PR descriptions, and code comments. Trigger it even when the user does not say "make this sound human" — if the output is prose someone will read as the user's own writing, run these checks before finalising. Do not trigger for code itself, structured data output, or short factual answers in conversation.
Three failure modes: sentences that would be true of any subject, sentences that are all the same shape, and over-correction of either.
Run the checks in this order: specificity, actors and naming, rhythm, performance, phrases, structures, punctuation, sourcing.
Look the word lists up; do not run them from memory. The two Cut sections name specific words and their replacements. A pass that recalls the rules instead of reading them will reproduce the exact substitutions the lists were built from; this has happened, with the offending word printed in the skill beside its fix. Search the draft for the listed terms.
Scope
A tells pass is subtractive. Remove, compress, or replace a vague phrase with a fact the draft already contains. Never add a claim, figure, example or paragraph the draft did not have; a rewrite that improves the argument is a different job, and the author did not ask for it. If the piece needs a fact it lacks, say so separately and leave the gap.
Never
- Invent a citation, statistic, quote or date
- Attach a real source to a claim it does not make
- Assert an unverifiable fact; flag the uncertainty or drop the claim
- Assert a claim wider than the cases actually examined. Five rows read and stated as a property of 23,486 is an overstatement even when every one of the five agrees. Qualify it (often, in the rows sampled) or narrow it.
- Attribute vaguely: some critics argue, many experts believe, observers have noted, studies show with no study. Name the source or cut it.
- Ship unfilled placeholders:
[insert X],[Your Name],[Company],[date] - Ship assistant framing: I hope this helps, Certainly!, Let me know if you'd like, Here is the, Would you like me to
- Ship process meta-commentary: I've structured this as..., This draft aims to...
- Leak literal
**bold**,##or-into a target that does not render it - Purge every em dash, however and list mechanically
- Strip specificity to dodge a flagged word
- Replace a flagged construction with a more polished one. The repair is plain, not clever: struggles at the point that matters most → is not fit for purpose, never → breaks on the reviews that matter most, the mixed ones. Two further constraints on repairs. A repair may not introduce a construction listed anywhere else in this skill; fixing a closing restatement with an em-dash clause trades one tell for another. And check the unit earns its place before repairing it, because sometimes the fix is deletion.
Cut: phrases
Flag clusters, not single uses. Cut the phrase or replace it with a fact.
- Puffery: stands as a testament to; plays a vital/pivotal role; underscores the importance of; a rich tapestry of; leaves a lasting legacy; is more than just; serves as a reminder; cements its place
- Editorialising: it is important/worth noting; it should be noted; no discussion would be complete without; critically; notably; importantly; arguably
- Tacked-on -ing clauses: highlighting the need for; reflecting a broader trend; underscoring its role in; paving the way for; showcasing. End on the fact.
- Manufactured contrast: not only X but also Y; it's not just X, it's Y; "That's not X. That's Y."
- Transition openers: Moreover; Furthermore; Additionally; In addition. However and But stay.
- Hedge stacking: may potentially; could possibly help to; might arguably suggest
- Elevated diction where a plain word works: arrives with → has; appended → added; the evaluation bar, the floor → the baseline; the pattern the EDA turned up → seen in the EDA. Figurative measurement nouns and phrasal verbs are the usual carriers. This is the single most repeated correction on record, so check it word by word rather than by feel.
- Engineering jargon in prose written for a non-engineer: contract → requirements, or just the fields; payload → the data sent; surface → the part people see; primitive → the basic piece. Fluent to the writer, noise to the reader.
- Consulting jargon: navigate (a challenge) → handle; unpack → explain; lean into → accept; deep dive → analysis; double down → commit; circle back → revisit; take a step back → reconsider; moving forward → next; on the same page → agreed; game-changer → name the change
- Vocabulary: delve, tapestry, pivotal, meticulous, vibrant, landscape, underscore, bolster, garner, boasts, realm, spine, leverage, seamless, robust, holistic
Check for the behaviour, not the string. Banning delve produces explore in depth.
Cut: structures
-
Say it once. Each fact appears in one place, counting the code, the figure and the table as places it can already have been said. The most mechanical form is an intensifier on a claim whose evidence follows: 21 exact duplicates above every field matches, including the review text; the bare prompt; moves the prose fields hardest above the two largest gains recorded anywhere. Also cut a sub-bullet that only re-expresses its claim, a clause in the claim that the evidence below demonstrates, and prose restating what the code above it plainly does.
-
Padding that defends or navigates. One test: does this unit report a finding? If it only protects the work or steers the reader, cut it.
- navigation: Out of scope here. Worth one line in the recommendations.; the exact baseline is quantified in the cell at the end of this section
- defending a choice: that bareness is deliberate, because…; a stored misspelling; the rule is not wrong in itself
- commentary on machinery the reader can see: constructing the client makes no network call; this cell is that check, and it costs four requests
- the consequence gloss the evidence already implies: reading
models.list()and assuming access would have picked a model that fails part-way through a paid run - the caveat the document later resolves: no noise bound has been measured, two sections before it is measured
- the paragraph whose subject is the document: they are set down separately from the recommendations because they apply to the next task rather than to this one
- the sentence that announces what the following sentences will do: Two further points hold the scale together.; Six prompt versions on one fixed sample produced findings that outlast this dataset. Delete it and start with the first point.
Two bounds. A signpost may ride inside a sentence that carries a finding (the class imbalance that question 3 quantifies); it may not occupy a unit of its own. And keep a reference to a requirement when the requirement is the reason for a choice (the brief asks for urgency, so a sixth field is generated); cut it when it is attribution for a fact that stands without it.
-
Negation-then-assertion: doesn't/isn't X; it's Y. Three tests, in this order.
- Entailment. If the asserted term already rules out the negated one, cut it whatever the reader assumed: a hypothesis, not a conclusion; stated, not silent. Decidable from the sentence alone.
- Live prior. Keep it only where the reader arrives holding the negated belief. This is not a prediction problem in the usual shape survives, because a data science reader assumes it is.
- Proportionality. A real prior still has to be worth the words. Stored in the source system rather than introduced here defends a typo's provenance against a challenge nobody made, and goes.
Name the concrete subject either way: Standing still doesn't close that gap; it leaves it open for a recognition-led competitor to go after directly → Standing still leaves ASB open for a recognition-led competitor to move first.
-
Tricolons where the third item is padding
-
The consequence chain: three or more harms in sequence, each entailed by the one before and none carrying a number (frustration, a spoiled shopping experience, reduced trust in the brand, fewer repeat purchases). Keep the one that costs money.
-
Examples that illustrate something other than the claim they sit under. Naming three legacy model ids illustrates legacy; it does not evidence half the catalogue is irrelevant. The test is whether the example evidences this claim, not whether it is concrete.
-
The synthesis closer: a final paragraph restating the piece, often opening In summary, Overall, Ultimately, Taken together
-
The bolted-on "Despite its success, X faces challenges..." section closing on a vague outlook
-
Repeated section shape: the same paragraph count under every heading
Performance tells
Once is fine. Three in a page is a tell.
- Fragments for emphasis
- Single-sentence paragraphs used as punctuation. One that carries its own claim and breaks up a long block is structure, not performance.
- Aphoristic closers: That's the whole game; That's the difference. Two exemptions. A sentence stating a criterion, threshold or decision rule is a claim and stays. And a maxim that generalises past the document is not a tell in itself, it is a tell in the wrong place: cut it from the middle of an evidence list, keep it where the piece gathers its general findings.
- The setup colon, where the left side exists only to promise the right: One problem: nobody shipped it; The refusals have a shape: the proxy admits floating aliases; this run measures by how much: 0.023. Test whether the left side carries information on its own. A label naming a category stays (The trade-off: this sample does not mirror the corpus; Highly rated: love is the single largest word), and so does a left side that is a claim in itself (The explanations converge on one mechanism: customers separate personal outcome from product endorsement).
- Rhetorical questions as transitions
- False-intimacy openers: Here's the thing; Let's be honest; The reality is. Same for any here's what / here's why / here's this construction; it announces the point instead of making it.
- Imperative closers: Start there; Do that first
- Withholding the point so it lands harder
- Personifying a method or dramatising sequence as obligation: the question Few-Shot now has to answer
- A metaphor where a concrete example would do more work
Test each short sentence: does it carry a hard claim, or is it a beat? Apply the test, not the pattern. These constructions are tells by frequency; every individual one that carries a claim survives.
Punctuation
The em dash is the most persistent tell in generated prose, and naming it is not enough to stop it. Reach for the mark that fits the job.
The last row is the job only the dash does: a final beat the sentence lands on. Everywhere else one of the other three marks fits better, and reaching for them is what stops the scatter. A spaced hyphen is not a substitute for any of them.
Formatting
- Match the house convention on heading case
- Use lists only for enumerable content, and always for it. Six ranked scores read better as a list than as a sentence, and so does a set of independent design decisions; enumerable does not mean short, and a bullet may carry a claim plus a sentence explaining it. Where a reader is scanning to check a point was covered, claim-led bullets answer that in one glance where a paragraph makes them hunt.
- No stray curly quotes or emoji where the document uses neither
- In prose, bold rarely and don't run
- **Term** — explanationdown a long list. In reference docs and specs it's fine.
Specificity
Ask of each sentence: would it still be true with a different subject in it? If yes, replace it with a fact.
Sharper form: if nobody could disagree with the sentence, it isn't saying anything. This catches summary sentences that restate the paragraph, transitions that only announce a topic change, and "factors influencing X" constructions that list categories instead of making a claim.
Actors and naming
Name who acts. Generated prose defaults to abstractions and passives in the subject slot, which is most of what makes a sentence true-of-anyone.
- Give the sentence its named actor where one exists: Protecting reputation through a peak → Protecting ChicStyle's reputation through a peak; as emphasis or sentiment → or customer sentiment; the client → the OpenAI client.
- Agentless passives hide a decision and the person who made it: it was decided, the approach was chosen, concerns were raised.
- Process nouns as subjects (the implementation, the adoption, the rollout) usually mean a person or team did something. Say which.
- Call a thing by the same name each time, and use a pronoun only when its referent is a noun the reader can point to. The artefact, for something the previous line called a
?token, is elegant variation; tolerate that and explains it directly point at whole clauses. All three make the reader re-read.
Rhythm
- Length: generated prose clusters at 18–25 words per sentence. If every sentence in a paragraph sits within a few words of the others, recast one long and one short; aim for a spread like 11 / 34 / 33 / 4 across four sentences, not four sentences all near 20.
- Openings: four consecutive sentences opening on a noun phrase. Rebuild one on a subordinate clause, a prepositional phrase, or a conjunction.
- Clause shape: watch runs of subject–verb–object declaratives with nothing embedded, and repeated "X, but Y" symmetry. Do not read this as licence for clutter; a subject carrying a relative clause and a mid-sentence appositive before its verb is harder to read than the plain sentence it replaced.
- Units: the boundary follows the thought. Merge sentences that are fragments of one point, split a unit carrying two. This runs both ways: running prose usually needs splitting, bullet lists usually need merging.
- Paragraphs: uniform three-to-four-sentence blocks, each opening on a topic sentence and closing on a summary sentence. Paragraph length is the strongest lever here; splitting a long block, including into a one-sentence paragraph carrying its own claim, does more than resequencing clauses inside it.
Vary where the meaning shifts weight, never on a schedule. A short sentence earns its length by carrying a hard claim; if it doesn't, make it a clause.
Register
Universal: sourcing, placeholders, vague attribution, specificity.
Format-dependent:
- Decks, runbooks and PR descriptions are legitimately list-shaped
- A consulting deliverable needs a summary, at the top rather than the bottom
- A landing page carries promotion by design; a report does not
- Slack and PRs tolerate more signposting than an essay
- A document reporting on data and a document answering a general question need different registers. An observation reports what was found; an answer to "how does X work" states the general case and uses the findings as evidence.
Provenance
Round 3, revised 2026-08-30 against 103 logged edits across 25 sections of one live document. The evidence, including what was falsified, is in observed-edits.md beside this file. Rules earn a place at two independent sightings; a rule that caused harm is qualified at one.