idiolect

July 8, 2026 · View on GitHub

Humanizers remove the AI. idiolect installs a human.

62 fully-specified voice profiles — real psychologies, systematic flaws, platform habitats, few-shot exemplars — plus a deterministic AI-tell scanner, a blind semantic auditor, and a closed verify-revise loop. Write, humanize, audit, and mass-produce short-form posts that read as distinct, specific people, with receipts instead of vibes. Runs on Claude Code, OpenAI Codex, and opencode out of the box.

idiolect (n.) — the variety of a language unique to one individual. The thing this plugin manufactures, measures, and defends.

The difference, in one post

The same request — "announce a kiln drop: the dark-blue glaze finally came out right, a couple pieces cracked, shop update Thursday" — scored by the same deterministic scanner (idiolect scan; 0 = clean, higher = more AI tells).

Without idiolect — default assistant output · scan 45 / 100 — slop

✨ Big news, friends! ✨ After weeks of anticipation, our stunning dark blue glaze collection is finally here — and we couldn't be more thrilled to share it with you! 🌊

Each piece is lovingly handcrafted and fired to perfection, embodying the beautiful imperfection of the ceramic journey. Yes, a few pieces didn't quite survive the kiln — but that only makes the survivors more special. 💙

Whether you're a seasoned collector or just beginning your pottery journey, these one-of-a-kind treasures are sure to elevate your space and bring a touch of handmade magic into your everyday moments.

Don't miss out — the shop update drops this Thursday! 🛍️ Link in bio.

Scanner flags: em-dash density, "elevate your," "journey" ×2, "whether you're a…", a six-hashtag pile, emoji density. Human texture: none — no specifics, no time anchor, no first-person action.

With idiolect — voice Freja, ceramicist, Copenhagen · scan 0 / 100 — clean · conform PASS, in-range

33 went in. 27 came out whole.

I opened 114 at 7, before anyone else was in.

The dark blue behaved. People have waited for it since firing 112 and I could not promise it, the kiln decides and not me. This morning the color came right on every blue piece, even the two that cracked. I wrote a full line in the log for it. Most firings get only numbers.

The kiln took its share too. 2 to the bin, a mug and the tall bottle, cracked at the foot, photo 4. You cannot both blow and have flour in the mouth, as we say at home.

4 are seconds. A little glaze pooling on the inside, nothing that leaks. They go 40 percent off with the flaw in the photo, not hidden.

Shop update Thursday 19:00 CET. Link in bio. 8 mugs, 5 bowls, 3 bottles in the dark blue. Mugs 260 DKK, bowls 340, bottles 520. Shipping inside EU is 89, the carrier's number.

Stoneware, fired to 1240 degrees. I look forward to see them go Thursday.

Written by the engine, then verified: the new competence layer kept it in clay (no borrowed jargon), soul is on (the blue glaze earns the one long line in the log; the loss is stated flat), and the L1-Danish idiom is her systematic error profile, not decoration. The scanner credits exact quantities, a time anchor, and first-person actions as human texture. Nothing here "sounds human" — it scans clean and conforms to a declared fingerprint.

Five voices, one brief, all conform-checked →

Why this exists

The popular humanizer skills are subtractive: a pattern list that removes "delve," em dashes, and rule-of-three. Useful — and structurally insufficient. Everyone running the same subtraction converges on a new detectable house style (uniformly short sentences, aggressive casualness, zero em dashes), because removal can't produce identity. Meanwhile RLHF models homogenize by design — co-writing with them measurably flattens stylistic diversity (Padmakumar & He 2024), hardest on non-Western writers (CHI 2025).

idiolect's claim, from the research up (see docs/THEORY.md, 40+ sources, corrections included):

Human-reading text = Scrubbed × Voiced × Crafted × Varied — multiplicative, a zero anywhere zeroes the product. So the plugin ships all four layers: a tell scanner (scrub), a voice roster (voice), a craft floor distilled from a century of copy and rhetoric research (craft), and structure-rotation + overlap + distance machinery (variance).

The demo

Same brief — "promote a $15/mo booking tool" — through five voices (full posts + receipts):

author	scan	one line of it
Dale, HVAC owner, Ohio (facebook)	0	"Diane won. 31 of the last 40 bookings came in after 6pm."
Zoe, art student, Glasgow (x)	9	"my dms are now exclusively for chat and crimes"
Keisha, HR director, Charlotte (linkedin)	0	"Small tools keep beating platforms here."
Hank, corn farmer, Nebraska (x)	0	"47 pickups the first week and her phone stayed in the kitchen."
Mai, bánh mì truck, Houston (facebook)	0	"Worth it for the 11:30 line alone."
the AI control	84	"🚀 …revolutionize the way you manage appointments…"

Zero cross-post overlap. Every post conform-checked against its voice's declared parameters.

Install

Claude Code

/plugin marketplace add nagisanzenin/idiolect
/plugin install idiolect@idiolect

Codex — codex plugin marketplace add nagisanzenin/idiolect && codex plugin add idiolect@idiolect, then bash scripts/install-codex.sh. opencode — add { "skills": { "urls": ["github:nagisanzenin/idiolect"] } } to opencode.json, then bash scripts/install-opencode.sh. Details, differences, and fallbacks: INSTALL-OMNI.md.

Use it in plain English

Commands exist, but natural language is the interface:

"write a launch post for my screenshot app, somewhere between reddit and HN energy"
"this reads like ChatGPT, fix it" → before/after scanner scores
"does this look AI-written?" → line-anchored tell report + blind semantic audit
"give me 4 posts for the launch — use the nurse and the sneaker guy, pick the other two"
"make me a voice from these three newsletters"
"write it like ME" → the self voice, built from your own measured writing

Slash forms: /idiolect:write, humanize, audit, campaign, synthesize, self (on Codex: $write, …). Voices are referred to like people — "the HVAC guy" resolves; slugs are for receipts.

What's inside

62 voices (voices/): ages 22–71, twelve L2/dialect textures (Vietnamese, German, Russian, Brazilian, Polish, Japanese, French, Egyptian, Korean, Danish, Nigerian English, Indian English…), every major platform habitat, eleven systematic error classes. Each profile: biography that generates opinions, stance engine, executable style patterns, anti-patterns, error rules, opener/closer pools, three scanner-verified exemplars. All fictional composites; a hard dignity rule (register habits yes, eye-dialect mockery never, competence first).
The engine (scripts/idiolect.py, stdlib-only): scan (tiered tell lexicon with model-era tags + construction regexes + stylometrics, line-anchored, cluster-scored), conform (draft vs voice parameters), distance (roster anti-clone floor, enforced at 0.2), overlap (campaign dedup), pick, ledger (structure rotation), fingerprint, self (private local corpus of your own writing), validate, selftest (25 tests), doctor.
Three agents: a blind auditor (reads only the text — catches the semantic tells regex can't: no stakes, symmetric enthusiasm, hollow specificity), an isolated campaign writer (one voice per context, so N posts can't bleed into each other), a synthesizer (corpus → new validated voice).
The data (data/tells.json): the tell lexicon as data, era-tagged (gpt4-era vs gpt5-era vs model-family tells), locally extensible via tells.local.json, calibrated against a fixture bench (bench/) with stated pass criteria.
Your voice (/idiolect:self): on Claude Code a hook quietly keeps a local corpus of your prose-like prompts (15+ words, no code/commands; self off kills it, self clear erases it, nothing ever leaves your machine); add real samples for a truer profile; drift tells you when your style is shifting under AI influence.

How a draft gets made

Pick voice → check the structure ledger (no repeated skeletons) → draft on the craft floor (one job, exact numbers, honest hook, no wrap-up bow) → scan (target ≤15) → conform (PASS required) → blind audit for campaigns → ledger the structure → deliver with a receipt line:

receipt: scan 5/100 (clean) · conform PASS · structure price/story/med · ledger recorded
texture: invented "after 6pm" detail — verify or keep as color

Honesty rules (built into the tools, not the fine print)

Never fabricates verifiable claims (stats, testimonials, partnerships) — invented interior texture is allowed and always disclosed in a texture ledger.
N voices presenting your thing in N registers = ghostwriting, supported. N voices posing as N unrelated satisfied strangers = astroturfing; the campaign skill refuses the shape and offers the legitimate one. No posting automation is included, anywhere.
The scanner lints your drafts; it does not convict other people. Style detection false-positives hit non-native writers hardest (Liang et al. 2023) — the audit output says so.

Limits, stated plainly

Lexical tells decay (the lexicon is era-tagged data, not scripture). Polite astroturf bottoms out mid-band for ANY deterministic scanner — that's the blind auditor's jurisdiction. No "undetectable" claims, ever: the receipt says no known tells, in-voice by measurement, distinct by measurement, survived a hostile blind read — nothing more, and nothing less than that either.

Extend it

/idiolect:synthesize grows the roster (measured from a corpus or invented into a gap — new voices must pass the same validate + distance gates as the built-ins). Custom voices live in $IDIOLECT_HOME/voices/, safe from plugin updates. The spec is public: docs/VOICE-SPEC.md.

License

MIT. The voices are fictional composites; any resemblance to real persons is coincidental. What you post is yours, on your accounts, under each platform's rules.