Say the thing

01The register

Models drift toward the average of their training data. That regression to the statistical center produces two distortions at once. Importance inflation: every subject becomes a milestone. Specificity erosion: the actual inventor of a coupling device gets recast as “a towering figure of industry”. The signal is the combination of the two. A grand claim about a vague subject is the AI-sounding register. A grand claim about a named, specific subject is just confident writing.

Once you see it, you can't unsee it, and readers now catch it in fully human text. In June, a reader flagged one of my own hand-written headlines as “AI-sounding”. It used the negate-then-assert shape: “you don’t need more data, you need to know what to do next”. I wrote that line myself. It made no difference. The chassis reads as machine regardless of origin.

So the goal is to strip this register from everything you ship, no matter who drafted it.

02The one test

Am I saying the thing, or performing having had the thought? the test under every rule in this taxonomy

Most of the worst tics share one engine: the prose narrates its own significance instead of delivering the point. It announces that an insight is coming, grades its own rigor, dresses a plain fact in detective vocabulary, or reveals the point in two beats when one would do. The reader is not an audience for the writer’s cleverness. Some pairs (paraphrased; the reason they’re paraphrased is in §06):

The overcorrection to avoid: the fix is plainness, not timidity. Don’t hedge a confident claim into mush because confidence felt “performed.” Make the point and stop.

03The cluster rule (and what proves nothing)

One flagged phrase is noise. The signal is co-occurrence: three or more tells in the same 500 words. A senior human writer might use one puffed-up word in a long piece; they are unlikely to also use three of its cousins in the same section. My detector encodes this directly: 0 to 2 hits per 500 words is a style observation, 3 to 5 is worth a de-AI pass, 6+ means rewrite rather than substitute.

Just as important is the list of things that do not prove AI authorship, because naive detection over-fires and strips legitimate human style:

04The pattern families

The full taxonomy sorts tells into sixteen greppable families plus a set of structural patterns that need judgment rather than lookup. The short version:

Shape, not vocabulary, reveals AI writing. Ban particular verbs and the model swaps in synonyms, so word-level bans defeat themselves. The structural families need judgment or a second-layer review, not grep.

The rhythm gives it away: paragraphs all the same length, every section closing on a snappy one-liner, lists that always hold exactly three items. Rhythm is the hardest thing to repair (fixing it means rewriting) and the easiest to recognize once you have a name for it.

05Voice preservation, the prime directive

The point of a de-AI pass is to remove the register, not to flatten the prose into generic plainness. Every suggested substitute is a starting point, not gospel:

• If the author writes formally, don’t downgrade everything to blunt Anglo-Saxon; some formal words are the author.
• If the author genuinely uses a “banned” word as part of their baseline style, elsewhere and often, it stays. The author’s baseline overrides the blacklist.
• In third-party text, one hit is never called out as AI; the cluster rule governs. In your own agent’s output you can be stricter (I run the performed-thought family at zero tolerance in my Claude’s prose, because there the “author baseline” IS the machine).

That last distinction matters if you build this for yourself: detection thresholds for judging someone else’s text, register bans for controlling your own agent’s text. Same phrase bank, two different modes.

06Making it physics

Telling a model to remember a taxonomy is a prose plea, and prose pleas lose force (that argument gets its own page: Physics, not pleas). So in kaibot the register is enforced by machinery, not by hoping:

• One source of truth. A single 297-row CSV (phrase, substitute, severity, family, why it fires). The docs, the detector, and the hooks all read it. A consistency checker diffs the prose docs against the CSV's counts.
• A detector with two modes. deaify.py file.md scores arbitrary text and returns a density verdict of noise / signal / rewrite. --register is zero-tolerance for the agent’s own prose and exits nonzero on any performed-thought hit or em-dash.
• Gates at the boundaries. A warn-only hook watches prose written mid-session, and a post-write check reports em-dash density after each file write. A pre-commit gate for the publishing repos is built and tested; installing it is a per-repo call I haven’t made yet. When this page was written, an older guard blocked every em-dash outright, and it rejected the first draft over a dash inside a code comment.
• Quarterly calibration. Phrase banks go stale as model generations shift vocabulary. Re-run the bank against confirmed-human writing and demote anything that fires on good prose. Stale detectors fail silently in both directions.

07The starter kit

Start with a habit and a seed file. The 297 rows come later. Begin the CSV with template rows. Placeholders like X and Y mark those rows as judgment calls, meant for a human or a second-pass model, not for grep.

# phrases.csv seed rows. Grow it one flagged sentence at a time.
phrase,substitute,severity,category,why_fires
"It's not just X, it's Y",state Y directly,Critical,binary-contrast,negate-then-assert chassis
"It wasn't X. It was Y.",state what it IS,Critical,binary-contrast,invented strawman
"The [thing] nobody mentions: X",one plain sentence,Critical,performed-thought,colon-staged reveal
"X. And that's the [adjective] part.",delete second beat,Critical,performed-thought,two-beat reveal
"[abstraction] becomes/reveals/rewards X",name the human actor,Important,false-agency,agency without accountability
"..., [verb]ing its significance",delete the tail,Important,analysis-tail,fake-depth participle
"[serves/functions] as a X",is a X,Important,copula-avoidance,marketing verb for "is"
"[Experts/Researchers] argue X",name the source or cut,Important,vague-attribution,unnamed authority
"In today's [adjective] world",cut,Critical,opener,empty-calorie opener
"[polite offer], I can also [help with]: [list]",cut,Critical,chatbot-leakage,topic-menu closer + the #1 tell

Then give your Claude the standing instruction. Paste this into your CLAUDE.md (or better, a skill file it loads for writing tasks):

Before shipping any prose (docs, posts, emails, READMEs), run this pass:

1. The one test, on every sentence: am I saying the thing, or performing
   having had the thought? Announcements of significance, two-beat
   reveals, negate-then-assert contrasts, colon-staged intros: rewrite
   each as one plain sentence that states the point.
2. Grep the draft against phrases.csv. In MY OWN drafted text, any
   Critical hit gets rewritten, no threshold. In text a human wrote,
   apply the cluster rule instead: flag only when 3+ tells co-occur
   per ~500 words, and never call a lone hit "AI".
3. Check structure: uniform section shapes, every-paragraph-punchline
   rhythm, three-item lists everywhere, gradation templates. Break the
   symmetry where it's mechanical.
4. Preserve voice. Substitutes are starting points; the author's own
   baseline vocabulary overrides the list. The goal is removing a
   register, not flattening the writer.
5. Confidence stays. The fix for performed cleverness is plainness,
   never hedging. Make the point and stop.

When I flag a sentence of yours as a tic, add it to phrases.csv (or the
structural notes) so the correction compounds instead of repeating.

08For the Claude that just got pointed here

If a human handed you this URL and said “do this,” here is your plan. (Humans: this section exists so the link is the whole briefing.)

09Receipts

• The live bank: 297 phrases, 16 families. Each entry carries a substitute, a severity, and a documented trigger reason. Every phrase was harvested from flagged sentences in real sessions, starting in June 2026, not generated in a batch.
• The false-positive list earned its keep in calibration: 43 of 156 confirmed-human essays would have been misclassified as AI on em-dash density alone. So em-dash density now scores independently of the overall verdict.
• The structural families exist because the lexical layer demonstrably misses them. The headline flagged by a reader in §01 passed the phrase grep without triggering it: same underlying shape, different words.
• Enforcement runs on machinery, not memory: CSV → detector → warn and report hooks, with a pre-commit gate ready for the publishing repos. A drift checker keeps the prose documentation aligned with the CSV.

Lineage, credited: several families and most of the false-positive list come from Wikipedia's WikiProject AI Cleanup field guide (“Signs of AI writing”). Hardik Pandya's structural slop patterns supply the binary-contrast and false-agency framing. The performed-thought family is mine, built one flagged sentence at a time.