humanize: A Rule-Based Skill Pair for Rewriting AI Prose and Scoring Its Detection Signals
Best static AI text humanizer. Two research-grounded LLM-agnostic skills that make AI writing sound human and relatable. Nine levers, 50+ peer-reviewed sources, 2024-2026 detection literature.
At a glance
- What is it?
- harshaneel/humanize ships two static SKILL.md files that any LLM agent can load. One applies nine documented rewriting levers; the other scores text against nine AI-signal categories and returns a verdict with confidence.
- Who is it for?
- Adopt this if you already run an agent harness that reads a skills directory and you want a documented, zero-dependency rulebook for editing AI prose. Skip it if you need a trained classifier or a detector that survives adversarial paraphrase, because the README itself names the learned-classifier ceiling and points to cross-model paraphrase as the way past it.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 67 days ago.
- What is it written in?
- Mainly HTML, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The Problem: AI Prose Reads Flat, and Detectors Score It That Way
The repository frames its target narrowly. AI writing diverges from human writing in measurable ways, and the detection literature is the instrument that locates those divergences. The README states the project is grounded in 50+ peer-reviewed sources through April 2026 and that the detection research "is the measuring instrument, not the target." The claim is that closing the gaps the research identifies is what makes prose read as human, and that lower perplexity-based scores from tools like ZeroGPT and QuillBot are "a consequence, not the goal." That distinction matters for evaluation. Anyone installing this to defeat a specific detector is optimizing for the wrong variable. The intended user is someone producing AI-assisted text who wants the output to carry natural rhythm, concrete specifics and a voice, and who wants a second skill that can audit a passage for the same signals. Both skills are LLM-agnostic. The README lists Claude Code, Codex CLI, ChatGPT, Gemini, Cursor, Aider, OpenCode, Continue and Copilot as hosts, with the caveat that "the install paths differ; the skill content is identical."
Two SKILL.md Files, No Runtime, No Trained Model
The architecture is deliberately thin. Each skill is a single SKILL.md file, described as static and rule-based with zero runtime dependencies. There is no model training step, no API call, and no detector in the loop. The host LLM reads the rulebook and follows it. The humanize skill applies nine humanization levers drawn from the detection literature; the README summarizes its trigger phrases as "humanize this", "make this sound more human", "make this less robotic", and "write this like a person". The ai-check skill performs what the README calls forensic analysis, scoring nine signal categories, citing every flag with evidence, and returning a verdict plus confidence plus an estimate of the AI-edited fraction. Six advanced hybrid techniques are documented inside the humanize skill for high-stakes use, though the README does not enumerate them in the section text provided. The practical consequence of the static design is that quality depends on the host model's instruction-following. A weaker model reading the same SKILL.md will apply the levers less consistently than a stronger one, and nothing in the repository compensates for that.
Installation Paths Across Agent Harnesses
The README gives a one-command path that covers three skill directories at once. Cloning the repository and running ./install.sh all installs to ~/.claude/skills/, ~/.codex/skills/ and ~/.agents/skills/. Adding the --copy flag switches from symlinks to self-contained files. Per-host instructions follow the same shape. For Claude Code, the manual route is mkdir -p ~/.claude/skills followed by cp -R humanize/humanize humanize/ai-check ~/.claude/skills/. Codex CLI uses ~/.codex/skills/, and the README notes that ChatGPT desktop and several OpenAI agent harnesses read from ~/.agents/skills/. OpenCode uses ~/.config/opencode/skills/, and the README adds that OpenCode also scans ~/.claude/skills/ for compatibility, so one clone serves both tools. Claude.ai and Claude Desktop cannot read from disk. The README instructs users to open Settings, then Capabilities, then Skills, click Create skill, and upload humanize/SKILL.md, repeating for ai-check/SKILL.md, then toggling each skill on per conversation. For chat agents with no skills directory, the README says to paste the raw contents of SKILL.md into a new conversation prefaced with an instruction to apply it whenever humanizing text. The symlink default is worth noting: ./install.sh without --copy means a later git pull updates the installed skills automatically, which is convenient but also means an upstream edit to a lever definition reaches your agent without a deliberate upgrade step.
Nine Levers, Nine Signal Categories, and the Detection Taxonomy Behind Them
The README organizes the supporting research into a detection methodology taxonomy with four families: zero-shot, classifier, watermarking, and hybrid. It also names the properties researchers measure, listing perplexity, burstiness, stylometry and discourse. The nine humanization levers are described as the rules the skill applies, drawn from that literature, and the ai-check skill scores nine signal categories against the same body of work. The README supplies accuracy figures from published research rather than from its own runs, and it is explicit that the benchmark section covers 25 inputs scored by two independent scorers. That is the extent of the evaluation material in the supplied text: the README does not report the benchmark's outcome, so a reader cannot tell from the README alone how the 25 inputs scored. The honest framing in the repository is that surface rewriting moves both the human-read impression and the perplexity-based numbers, and that the same edits do both. Whether that holds across languages, domains or lengths is not something the material establishes.
The Learned-Classifier Ceiling the README Admits
This is the section that decides whether the project fits a given use case. The README contains a section titled what this rule-based approach cannot do, described as the learned-classifier ceiling and why it exists. It names Grammarly and GPTZero as examples of learned classifiers that a static rulebook does not defeat. Complementary techniques are listed as the way to close the gap: cross-model paraphrase, base-model rewriting, and manual edits. Read plainly, the project does not claim to beat trained detectors, and it says so in its own table of contents. The failure mode follows. If your workflow requires text to pass a specific commercial detector, a rule-based rewrite is the wrong instrument, because the detector is trained on distributions the rules do not model. A second constraint is that the skills are prose-oriented. Nothing in the material suggests they handle code, structured data, or short strings such as error messages, and applying a rhythm-and-voice rulebook to a JSON payload would be a category error. A third is host dependence: because the skill is instructions rather than code, two agents given the same input can produce different output, and there is no deterministic test to run against a rewrite.
Where It Sits Next to a Learned Detector
The natural comparison is a trained classifier such as GPTZero, which the README itself names. The difference in approach is structural rather than incremental. A learned detector ingests a passage and returns a probability derived from patterns it was trained on; it is a black box with a score. humanize goes the other direction. It is a rulebook that tells a model which properties to change, and ai-check is a second rulebook that reports which of nine signal categories a passage trips, with cited evidence for each flag. One produces a number; the other produces an editable list of reasons. That makes ai-check more useful during revision, because a writer can act on a named flag, and less useful as a gate, because a rule-based verdict is not calibrated against a labeled corpus the way a trained classifier is. The README's complementary-techniques section implicitly concedes the division of labor: rule-based rewriting handles the surface, and cross-model paraphrase or base-model rewriting handles what the rules cannot reach. Choosing between them is not a matter of which is stronger. It is a matter of whether you need a reason to edit or a score to report.
Maintenance, Updates, and the MIT Licence
The repository is MIT-licensed and not archived, with the last push dated 2026-07-10. No releases were retrieved, so there is no versioned artifact to pin; the unit of distribution is the repository itself. Updating is a git pull in the clone directory, and because the default install uses symlinks, that pull propagates to ~/.claude/skills/, ~/.codex/skills/ and ~/.agents/skills/ without a reinstall. Uninstalling and contributing have their own README sections, though the supplied text does not detail their contents. The maintenance cost that matters here is editorial rather than technical. The skills cite research through April 2026 and the README describes the detection literature as an active field, so lever definitions can drift as the underlying papers are superseded. Nothing in the repository pins a source list to a version, which means an upstream edit to SKILL.md changes agent behavior on the next pull. MIT terms permit commercial use and modification; the README also has a Citation section, which suggests the author expects academic or research reuse. That is a description of the licence text, not legal advice, and anyone redistributing a modified SKILL.md should read the LICENSE file directly.
Editorial conclusion
Adopt this if you already run an agent harness that reads a skills directory and you want a documented, zero-dependency rulebook for editing AI prose. Skip it if you need a trained classifier or a detector that survives adversarial paraphrase, because the README itself names the learned-classifier ceiling and points to cross-model paraphrase as the way past it. Before installing, open humanize/SKILL.md and ai-check/SKILL.md and confirm the nine lever definitions match the style guide you actually enforce.
Community notes