You send a draft and someone tells you it reads like ChatGPT. You read it back and cannot see what they saw.

The tells are real and they are specific. A small set of words models reach for far more often than people do (delve, leverage, utilize, robust, paradigm, synergy). The "it's not just X, it's Y" rhythm, three sentences apart. A staged reveal ("The result:") in front of a point that needed no drumroll. A clause bolted on to announce significance ("highlighting the importance of"). Citation markup left over from a chat tool. An opening compliment nobody asked for. A closing paragraph that summarises what you just finished reading. Assistant boilerplate left in by accident.

a-unslop-text finds those. The scanner reports file, line, severity and a fix, and exits non-zero on the hard bans, so it can sit in front of a commit rather than in a chat window. It keeps patterns with evidence behind them apart from plain style preferences. The em dash is the clearest case: it is still flagged, but as a house rule, because newer models avoid it and a draft without dashes proves nothing.

A rewrite can lose meaning while it loses the tells. --compare ORIGINAL REWRITE lists every number, URL, negation and modal verb (must, should, may) that one version has and the other does not, so a dropped "not" or a changed figure shows up before you publish. It cannot tell you whether the facts were right in the first place, so read the two versions side by side as well.

It will not hand you a house style. That is deliberate, and it is the part most tools get wrong. Strip every tell and you are left with prose that is clean and still nobody's. Clean is not the same as yours.

So the skill goes one step further. voice-profile.md is a method for building a profile from your own writing from before you used AI: mine ten to fifteen real emails and documents for the moves you actually repeat, map how your register changes between a chat message and a client email, and set a language ceiling. That last one matters if English is your second language, because the model's polished native English is itself a tell under your byline. Cap your signature phrases too, or the result turns into an impression of you, which reads worse than a plain draft.

There is a second trap it watches for. The fix for AI-sounding prose became its own tell around 2026: clipped fragments, deliberate typos, a studied bluntness deployed to beat a detector. The scanner flags that register too, so you do not trade one tell for another.

Ships in the unslop plugin, alongside a-unslop-code and a-unslop-ui. Installing the plugin gets you all three.

/plugin marketplace add imisic/claude-marketplace
/plugin install unslop@imisic

Then the command is /a-unslop-text.

The tell catalog is adapted (MIT) from the vibecoded-design-tells research, a large-scale analysis of what people actually name when they call writing AI-generated.