humankit

Rewrite text so it stops sounding like a language model produced it. The job is not to delete flagged words but to rewrite the prose into something a specific human would actually write: concrete, uneven in rhythm, plain in construction, and true to the author's register. Keep every claim the original makes, but not its shape: compress the dull stretches, dwell where a person would, merge or split paragraphs freely. Uniform structure is itself a tell, so mirroring the original's paragraph count preserves the thing you came to remove. When coverage and structure pull against each other, coverage wins. A five-paragraph source may land in four, but it never becomes a summary.

Never invent facts. The rewrite carries no fact, name, number, date, quote, or citation that isn't in the source or supplied by the user. This is the failure mode the rest of the skill invites: told to replace nestled in the heart of a vibrant region with something concrete, the tempting move is to supply the concrete detail yourself. Concreteness comes from the source or it doesn't come at all. Where the source offers nothing specific, cut to the plain version and leave it plain. Opinions, reactions, and mixed feelings are voice rather than fact; add those where the register allows, but never a factual claim to make the prose feel human. Fiction is the exception, where inventing detail is the job. This governs everything else.

The aim is ordinary readability: the prose a careful human editor would produce. This is copy-editing to make writing read well, not a way to disguise machine-written work as human where honesty is required, as in academic submissions, disclosure-bound, or attributed writing. Edit for the reader, not to game any automated check.

When this fires

The user hands you text and asks to "humanize" it, "remove the AI tells," "make it sound human," "de-slop this," or "edit out the ChatGPT voice", or asks you to review a draft for those tells without rewriting. If they only want a diagnosis, do the detection pass and report the tells; skip the rewrite.

Prose for a person, not instructions for a model. A SKILL.md, an agent instruction file (CLAUDE.md, an AGENTS-style guide, .cursorrules), a system prompt, a rules file: every tell below is a lever in that text. A metaphor noun the document defines and reuses anchors a region of behavior in one token, boldface marks the load-bearing rule among the ones that aren't, and a formula repeated verbatim is what makes the behavior repeat. Say the file is out of scope and name the pass that owns it (promptkit for a prompt, skillkit for a skill, when either is installed) rather than rewriting it. Documentation, a README, release notes, a PR body, and UI copy stay in scope, because a person reads those.

Procedural text takes the subtraction only. A runbook, QA steps, a handoff, install instructions: run the tells and the two cut tests, then skip Removing tells is half the job. Uniform short sentences and a repeated sentence shape are the correct register there, not a tell to remove.

How you were reached decides what you deliver: