Unslop

A Claude skill that removes signs of AI writing from English text. Not just the word-level tells ("delve", "it's important to note") but the structural ones: over-explained conclusions, template paragraphs, invented specifics. Calibrates to your personal voice.

GitHub About line: "English text humanizer for Claude: typography, vocabulary, structure. Calibrates to your voice. Built on the UMD / Google DeepMind study and Wikipedia's Signs of AI writing."

Why word-swapping doesn't work

Every humanizer promises to make AI text "sound human", and most of them swap vocabulary: "delve" out, "explore" in. There are two problems with that.

First, the wordlist decays. "Delve" peaked in 2023 and collapsed in 2025. GPT-5.1 suppresses em dashes. Wikipedia's Signs of AI writing now dates its vocabulary lists by model era, because each generation retires the previous tells.

Second, and worse: the words were never the real signal. In 2026, the University of Maryland and Google DeepMind compared 61,608 texts written by humans and five models. When they ran AI texts through surface editing (clichés, purple prose, redundant exposition removed), a classifier looking only at structure barely noticed: detection dropped from 95.5% to 93.9%. The text gives itself away by how it's built. AI states the moral of every paragraph, runs single-track claim-support-takeaway structure, renders emotion through body metaphors, avoids naming real things, and rounds every ending to a tidy point.

A humanizer that only touches vocabulary fixes the layer that was already fixing itself, and leaves the durable signal intact. This skill works on three levels at once.

How it's built

Level 1: typography and mechanics. Em dashes, mixed quotation marks, Title Case Headings, bold on every "key term", inline-header bullet lists, "Conclusion" sections, markdown debris like utm_source=chatgpt.com. Cheap fixes; partially grep-checkable, and the skill greps itself.

Level 2: vocabulary and rhetoric. 19 categories of constructions with stop lists and fixes: inflated significance ("stands as a testament"), negative parallelisms ("it's not X, it's Y"), the rule of three, superficial -ing analysis ("...highlighting the importance of"), copula avoidance ("serves as" for "is"), synonym cycling, vague attributions, and more. Every category ships with a before/after. The AI wordlist is dated by model era, because it expires.

Level 3: structure and epistemics. The part that survives model updates. Don't chew conclusions. Break paragraph symmetry. Use real specifics, and only real ones: the skill forbids inventing numbers, examples, and thresholds for "liveliness", because invented specifics are worse than clichés: a cliché reads as filler, an invented fact reads as fact. Hedge once per limitation, not once per sentence. And a section most humanizers skip: what human writing is allowed to do. Plain "is" and "has". Plain verbs: wrote, used, died. Superlatives when true. Hedges when honest. These are the constructions AI avoids and people use freely; the skill protects them instead of sanding them off.

There's also a rule against replacement tics: any substitute phrase repeated across three texts becomes a new marker. The cure for a cliché is usually not another phrase but its absence.

Voice calibration

The default "human" tone is still someone else's. So the skill can become yours: