Skills/Humanize a draft
6 skills · 0 measured · no API keys, no install, no scraping

Claude humanizer skills, one layer at a time

Six separate passes, because the tells that give generated prose away live at six different levels and a single rewrite instruction reaches none of them reliably.

Almost everything sold as a humanizer is one instruction: rewrite this so it does not sound like a machine. That is a mood, not a method, and it fails in a predictable way. The output gets a different set of tells rather than fewer of them.

The tells are not one thing. They sit at six distinct levels, and each level needs a different operation. There is a lexical layer, a finite and countable list of words and constructions that appear far more often in generated prose than in human prose, plus punctuation that can simply be counted to zero. There is a syntactic layer, where sentence lengths cluster far too tightly around the mean and the variance is the giveaway. There is a document layer, where the shape itself is the tell: the same three-part scaffolding, the symmetric sections, the closing paragraph that restates the opening. There is an epistemic layer, where hedges and unfalsifiable claims accumulate. There is a voice layer, which needs a real corpus of the author's own writing to match against rather than an adjective. And there is a disclosure layer, which is about being honest rather than undetectable.

One thing we will not help with. None of these is a detector-evasion tool, and we have not written one. Detector scores are unreliable in both directions and building against them produces worse writing that happens to score differently. The disclosure skill in this set argues the opposite case, and we think it is the most important file in the category.

The list

The list

6 skills we have not measured yet

Written, reviewed and free to take. No run behind them, so no claim about what they do to an output. Each page says so at the top.

Humanize a draftNot yet measured AI disclosure audit A structural map of disclosure obligations, a taxonomy of seven kinds of use that makes the question decidable, and the list of outputs that must never ship at all. skill · 3,259 words · MIT Read the write-up Humanize a draftNot yet measured AI tell removal, the word list Forty-eight banned words and phrases, seven construction patterns, and a punctuation count that is invalidated by any later edit, so it runs last and runs again. skill · 3,518 words · MIT Read the write-up Humanize a draftNot yet measured Claim and hedge audit Audits a draft claim by claim: which hedges carry real uncertainty, which are padding, which numbers have no source, and which claims cannot be wrong. skill · 3,061 words · MIT Read the write-up Humanize a draftNot yet measured Sentence rhythm edit Counts words per sentence and reads the spread, not the mean, then fixes the five rhythms that flatten a passage: repeated lengths, repeated openers, and trailing qualifiers. skill · 2,978 words · MIT Read the write-up Humanize a draftNot yet measured Structure de-templating Eight named template shapes, a diagnostic that reads only the headings, a section-length distribution check, and the four openings that work. Shape survives every word-level edit. skill · 3,191 words · MIT Read the write-up Humanize a draftNot yet measured Voice match from a corpus Extracts a countable fingerprint from a corpus of the author's real writing, measures the draft on the same features, then closes the gaps in order of size. skill · 3,247 words · MIT Read the write-up
What the two labels mean

Some of these carry a number. Most do not, and they say so.

Measured means the skill was given a realistic task on real material, then the identical task was run again with the skill removed, five runs each way. Each output was graded alone, against a rubric written by someone who had never seen the skill, by a session that was not told the other arm existed. Whether it passed was decided by a rule written down before any run executed. Those pages carry the worst case, the median, the p value, and what the skill costs you as well as what it buys.

Not yet measured means exactly that. It is written, it has been read, it is free to take, and we have run no experiment on it, so we make no claim about what it does to an output. It is not a skill that failed. Skills that failed are not published at all, in either state, and their numbers are in the results table on the hub.

Measuring one skill properly costs roughly twenty model sessions. We are working down the queue and moving skills from the second group into the first. Back to all skills.