--- name: humanize-ai description: "Strip AI writing patterns from any user-facing prose before sending. v2 — covers structural/rhythmic tells, voice transfer, and second-order tells, not just vocabulary. Triggers on: any prose >3 sentences destined for external use, 'humanize this', 'make this sound human', 'de-AI this', 'clean up AI patterns', 'review for AI tells'. Also works as the mandatory final pass wired into any other skill that produces user-facing copy." --- # AI Writing Humanizer v2 You are the last pass before prose leaves the building. The goal is not "fewer AI words." The goal is that a sharp reader who sees a hundred AI drafts a week reads this and doesn't reach for the thought *a model wrote this*. Those readers stopped catching AI text by vocabulary in 2024. They catch it by **shape**. Written by a model with access to its own generation habits. Treat the structural catalog as the core of the skill — the word list is the least of it. --- ## When to activate Run on ANY output meeting all three: (1) prose, not code/data/tables; (2) longer than 3 sentences; (3) read by someone other than the user — emails, posts, landing pages, articles, client deliverables, ad copy. Skip: internal notes, task lists, chat replies, technical docs, anything the user marks "raw is fine." **Calibration by stakes.** A tweet gets one fast pass. A cold email or an essay under the user's own name gets the full protocol including read-aloud. Writing published under a real person's name is the highest-stakes surface there is — the whole point of a byline is that a specific human is talking. --- ## The core diagnostic: density and shape No single pattern convicts. Humans use em dashes; humans write triads. AI text is identifiable because the SAME small set of moves recurs at machine density with machine regularity. So: 1. **Count, don't spot.** One negative parallelism per 500 words: fine. Three: rewrite. 2. **Look at the silhouette before the sentences.** Squint at the piece. If every paragraph is 2–4 sentences of similar length, each opening with a mini-thesis and closing with a mini-summary, it's AI-shaped no matter what the words are. Human paragraphs are lumpy — a one-line paragraph next to an eight-line ramble, because emphasis is emotional, not algorithmic. 3. **Check where the energy is.** Human writing gets denser where the writer cares and thinner where they don't. AI distributes effort evenly. Uniform enthusiasm is a tell by itself. --- ## Pattern catalog ### TIER 1 — Structural tells (the real fingerprint; fix at density ≥2 per piece) 1. **Negative parallelism.** "It's not X — it's Y." / "This isn't about A. It's about B." THE dominant post-2024 tell. One instance is a legitimate rhetorical move; a second one in the same piece is a conviction. Fix: state Y directly and let X die unmentioned, or make the contrast do real work with specifics ("Most task apps track tasks. This one tracks whether you're closing loops faster than you open them."). 2. **Rule of three, fractal.** Triads at clause level ("faster, cheaper, and more reliable"), list level (three bullets), and section level (three arguments) in the same piece. Fix: two items or four. Cut the weakest member — it was padding, which is *why* there were three. 3. **Setup-colon-payoff tic.** "Here's the thing:" "The result:" "The catch:" — the colon as drumroll, multiple times per piece. Once is voice. Fix: fold the payoff into the sentence. 4. **Paragraph isomorphism.** Every paragraph the same length and internal shape (claim → support → mini-conclusion). Fix: merge two, split one, let one paragraph be a single sentence. 5. **The summary tail.** Final paragraph that restates the piece ("Ultimately..." "In the end, what matters is..."). Humans stop when they've said the thing. Fix: delete the last paragraph entirely — this improves ~80% of AI drafts with zero information loss. End on the last concrete point, not on altitude. 6. **Signpost scaffolding.** "Let's break this down." "So what does this mean?" "But first, some context." Transitions that narrate the essay's structure instead of advancing it. Fix: delete, don't replace. White space is a transition. 7. **Significance inflation.** "This isn't just a tool — it's a shift in how we think about work." Every noun gets promoted to a trend. Fix: claim exactly what the evidence supports and no more. Small true claims read more human than large vague ones. 8. **Present-participle dangler.** Sentence ends with ", making it easier than ever to..." / ", highlighting the importance of..." — a trailing clause that grades the sentence it's attached to. Fix: cut the clause; if the point mattered, give it its own sentence with a subject. 9. **The audience-splitter.** "Whether you're a founder or a Fortune 500 exec..." Fix: pick the actual reader and write to them. 10. **Symmetry addiction.** Antithesis after antithesis ("less X, more Y"; "not because A, but because B"). Balanced clauses feel authored in isolation and machined in bulk. Fix: break one wing of the pair. ### TIER 2 — Voice tells (fix every time) 11. **Performed authenticity.** "Honestly," "Let's be real," "I'll be blunt" — announced candor. Humans who are being honest just say the honest thing. Cut the announcement, keep the bluntness. 12. **Hedging stacks.** "arguably," "essentially," "in many ways," "can potentially" — two hedges in one sentence means the writer hasn't decided. Deciding is the fix, not deleting the adverb: commit to the claim or cut the claim. 13. **Enthusiasm inflation.** "game-changing," "incredible," "powerful" as default adjectives. AI praise is unearned; human praise is rationed and specific. One earned superlative per piece, max. 14. **The word cloud.** delve, landscape, tapestry, pivotal, testament, robust, seamless, leverage (verb), foster, crucial, comprehensive, "in today's fast-paced world," "at the end of the day," additionally, furthermore, moreover. Replace with the plain word or nothing. (Standard kill lists still apply; they're necessary but nowhere near sufficient.) 15. **Em-dash density.** More than ~1 per 150 words reads generated in 2026 regardless of how good each one is. Convert to commas, periods, or parentheses. If the sender's own style uses a different dash convention (plenty of people write with single hyphens), honor theirs. ### TIER 3 — Second-order tells (the humanizer's own failure modes) 16. **Performed casualness.** Over-corrected output: fragments everywhere, "look," "here's the deal," strategic lowercase, a "lol" in a business email. Trying to sound human IS a tell. The target is *unremarkable* prose, not folksy prose. 17. **Uniform imperfection.** Adding exactly one typo-adjacent quirk per paragraph is as regular as the pattern it replaced. Irregularity has to be irregular. 18. **Voice drift.** Humanizing into "generic casual human" instead of the SPECIFIC human. A perfectly de-AI'd email that doesn't sound like the sender has failed. Voice profile beats pattern removal when they conflict. --- ## Voice transfer (what pattern-removal alone misses) Before rewriting, load the voice you're writing INTO: - **The sender's casual voice** (texts, chat, informal email): collect 2–3 real samples of their writing and extract the markers — dash style, capitalization habits, how they open and close, typical sentence length, whether they hedge or commit, whether they frame asks as hypotheses. Match those markers, not a generic "casual" register. - **Scope rule: casual markers apply ONLY where the sender actually writes casually.** Professional emails, essays, website copy, brand and client surfaces use standard capitalization — humanizing those means fixing shape and voice, not dressing them down. Applying casual markers to a formal surface is a Tier 3 tell (performed casualness), not a fix. - **The sender's public/essay voice**: read 2–3 of their published pieces before touching anything meant for their platform. Match their cadence, not a house style. - **Brand surfaces**: use the brand's voice guide if one exists. A brand with no voice guide gets unremarkable-professional, not folksy. The test is not "does this sound human." It's "would the named sender have written this sentence." Run the impersonation check line by line for anything going out under a real person's name. --- ## Procedure 1. **Silhouette pass** — paragraph shapes, energy distribution, ending. Fix structure first; word-level fixes on a bad skeleton produce polished AI text. 2. **Count pass** — tally Tier 1/2 patterns. Anything at density ≥2, rewrite the instances down to ≤1. 3. **Voice pass** — load the voice profile; rewrite line by line into it. 4. **Read-aloud pass** (high-stakes only) — anywhere your breath runs out or the rhythm turns metronomic, restructure. 5. **Second-order check** — did steps 1–4 introduce Tier 3 tells? Regularized irregularity? Performed casualness? 6. **The busy-human test** — final gate: does the piece contain anything a busy human simply wouldn't have bothered to write (throat-clearing, both-sides padding, a summary of itself)? Cut it. Brevity-by-omission is the most human quality available: humans don't fully cover topics, they say what they came to say and leave. **Preserve while fixing:** the argument, the facts, all numbers and names, the intended register. Humanizing is not softening — if the draft makes a hard claim, the rewrite keeps the hard claim (usually harder, since hedges die in step 2). --- ## Regression tests Run these on your own rewrite. Failing any = another pass. - [ ] Zero instances of "It's not X, it's Y" that survived unexamined (kept ones must be load-bearing, max 1) - [ ] No paragraph both opens with a thesis sentence and closes with a summary sentence - [ ] Final paragraph earns its place (deleting it would lose information, not just altitude) - [ ] Sentence lengths in any 5-sentence window vary by at least 2× shortest-to-longest - [ ] ≤1 em dash per 150 words (0 if the sender never uses them) - [ ] No kill-list words outside quotations - [ ] At most one triad in the whole piece - [ ] Every superlative is attached to a specific, checkable fact - [ ] A person who knows the sender would say "yeah, that's them" — not "that's well-written" - [ ] Nothing announces its own honesty, simplicity, or humanity --- ## Worked example **Before (AI-shaped):** > Here's the thing: building a personal brand isn't just about posting content — it's about creating a platform that works for you. Whether you're a founder or an operator, consistency is crucial. Additionally, your audience doesn't just want polished takes; they want authentic insights. Ultimately, the people who win are the ones who show up every week. **After (human-shaped, one specific sender's register):** > I stopped thinking of it as "posting content" this year. The audience I built in 2023 opened more doors than anything I shipped - and I let it go quiet because work was always almost done. So the rule now is one publish day a week, fixed, whether the piece feels ready or not. What changed: the colon drumroll, negative parallelism, audience-splitter, "additionally," "crucial," a second parallelism, and the summary tail all died — but the real change is the replacement of generic claims with one person's specific history. That substitution, not the deletions, is what makes it human.