humanizer

A standalone, pure-prompt skill that rewrites AI-sounding prose so it reads as
genuinely human, and rewrites in a specific writer's voice when a sample or
style profile is available. It also conservatively cleans suspicious invisible
Unicode in supplied prose. It is just instructions: SKILL.md plus a few
reference files. No scripts, no dependencies, no network access.

Where this comes from

This skill was built from the voice-preservation logic that powers
Scriveno (formerly Scriven), an
AI-native longform writing system whose core promise is narrow and
high-stakes: drafted prose should sound like the writer, not like AI.
Scriveno is a spec-driven creative-writing, publishing, and translation
pipeline for AI coding agents. It profiles a writer's voice, loads that
Voice DNA into every drafting step, keeps each unit on fresh context, and
runs a Polish pass (editor review, line and copy edit, voice check,
originality check) so prose stays specific to the project instead of
collapsing into generic machine output.

humanizer lifts the part of that pipeline that finds and removes AI writing
tells while protecting a writer's authentic voice, and packages it as a
standalone, tool-agnostic skill: the same de-slop, restraint, and
voice-matching philosophy Scriveno applies across a full manuscript, usable
on any prose in any supported tool. If you want the whole writing,
publishing, and translation pipeline rather than just this de-slop layer, see
Scriveno: https://github.com/hannsxpeter/scriveno (npm package: scriveno-cli).

Why this one is different

• Faithfulness over liveliness. A three-layer guard plus a mandatory
  meaning check stop the most common humanizer failure: inventing plausible
  facts, causes, or quotes to make a rewrite "sound concrete."
• Restraint by design. A density pre-check scales effort to evidence, and
  an explicit do-not-flag reference protects already-human writing instead of
  laundering it into smooth, average prose.
• Variance, not synonym-swapping. It targets the real signature of
  machine text (uniform rhythm) rather than relocating it with a thesaurus.
• Voice-first. When a writing sample or a VOICE.md / STYLE-GUIDE.md
  is present, it rewrites in that author's cadence and stance, not generic
  neutral prose.
• Opt-in edge. A stance mode adds opinion and punch on explicit request,
  hard-blocked from inventing content.
• Conservative text hygiene. It removes high-confidence invisible
  formatting residue while preserving script joiners, direction controls,
  display selectors, locale spacing, code, and exact-value spans when they are
  load-bearing.

What it removes

A 32-pattern catalog in six families, each with detect / why / before-after /
restraint notes: inflated significance, promotional and evasive language,
formulaic structure, lexical tics, syntactic tics, and formatting artifacts,
including chat-UI contamination, debunking-pose headings, and diff-anchored
writing.

Prompt-only text hygiene

Every run now includes a text-hygiene preflight. When the host exposes the
characters, the skill can remove stray zero-width spaces, soft formatting
controls, unexpected direction controls, free-floating tag characters, and
copy-paste spacing residue. It reports what changed and preserves ambiguous or
load-bearing Unicode instead of normalizing it blindly.

This feature adapts the prompt-compatible text-layer ideas from
watermarks-remover.
That project also offers scripts for C2PA, EXIF, XMP, document metadata, and
media processing. Humanizer remains pure prompt, so those container and media
features are deliberately outside its scope. A prose rewrite may disturb
statistical token patterns as a side effect, but this skill cannot verify or
promise their removal.

Supported tools

Every adapter points the agent at the same SKILL.md and references/, so
the workflow is identical across tools. The simplest universal install is to
clone this repository into (or alongside) the project you are writing in.

Usage

Ask, in plain language, to humanize, de-slop, or fix prose that "reads like
AI," "sounds corporate," or is "too generic," or to make a draft sound like
you or a named author. You do not need to say the word "humanize."

For voice-matched output, do one of:

• paste a sample of the target writing,
• name a well-known author, or
• keep a VOICE.md (schema in references/voice-matching.md) or a
  STYLE-GUIDE.md in the project; it is discovered automatically.

Every run returns the rewritten text plus a short report: what changed, what
was deliberately left alone, a text-hygiene status, a meaning check, and a
next step.

Scope

This skill improves prose quality and authentic voice. It is not designed or
tuned to defeat plagiarism checkers or AI-detection systems, and it names no
detector. Requests framed as passing AI work off as a person's own for a
graded or contractual assessment are reframed toward the quality-and-voice
use the skill actually serves. Text hygiene applies only to characters in the
supplied prose. It does not inspect or strip file-container provenance or
media marks.

Layout

SKILL.md                        orchestrator: workflow, guardrails, output contract
AGENTS.md                       cross-tool entry point (Codex, Antigravity, OpenCode, Pi Coder)
GEMINI.md                       Gemini CLI context
.cursor/rules/humanizer.mdc     Cursor project rule
.github/copilot-instructions.md GitHub Copilot instructions
.windsurfrules                  Windsurf rules
.clinerules                     Cline rules
.continue/rules/humanizer.md    Continue / Zed rule
CONVENTIONS.md                  Aider conventions
references/tell-patterns.md     the 32-pattern catalog (read in Pass 2)
references/do-not-flag.md       false positives, human markers, stop conditions
references/voice-matching.md    voice discovery and application
references/text-hygiene.md      conservative invisible-Unicode cleanup
references/examples.md          five worked end-to-end runs
evals/evals.json                verification cases (not part of the runtime skill)

License

MIT. See LICENSE.