harshaneel/humanize

humanize – Rule‑based “human‑like” writing skills for LLM agents

What it is – A tiny, MIT‑licensed repository that ships two static skills for any LLM‑based coding or writing assistant (Claude Code, Codex CLI, ChatGPT desktop, Gemini, Cursor, Aider, OpenCode, etc.).

Skill	Goal	How it works
humanize	Rewrites or generates text so it reads like a real person – natural rhythm, concrete details, a consistent voice.	A single SKILL.md file containing a rule‑book derived from nine “humanization levers” identified in the AI‑text‑detection literature (perplexity reduction, burstiness injection, hedge surgery, specificity insertion, voice matching, etc.). No model, no API calls – the host LLM simply follows the instructions.
ai-check	Forensic analysis of any passage for AI‑generation signals.	Another SKILL.md that scans the input for nine signal categories, cites every flag, and returns a verdict (Human / Likely Human / Uncertain / Likely AI / AI) plus confidence and an estimate of how much of the text was AI‑edited.

Installation – one command works for all supported agents

# Clone the repo (required for any agent that reads from disk)
git clone https://github.com/harshaneel/humanize.git
cd humanize

# Install to every supported skill directory (Claude, Codex, generic agents)
./install.sh all   # add --copy if you prefer copies instead of symlinks

Claude Code and Codex CLI also support a marketplace workflow:

claude plugin marketplace add harshaneel/humanize
claude plugin install humanize@humanize   # same for Codex

For web‑based Claude or any chat UI you can upload the two SKILL.md files via the Settings → Capabilities → Skills screen, or simply paste their contents into a conversation as a system prompt.

Using the skills

Humanize – rewrite text:

/humanize:humanize

[paste your AI‑generated text]

Or ask the model directly:

Please humanize this text: …

AI‑check – audit text:

/humanize:ai-check

[paste the passage]

Or:

Does this sound AI? …

Both commands work with the short names (/humanize and /ai-check) when the repo is cloned locally.

Voice‑matching (optional)

Provide a short personal writing sample before the target text; the skill extracts rhythm, vocabulary, and structural quirks (using the HyPerAlign method, arXiv 2505.00038) and applies them to the rewrite.

Benchmark reported in the README

• Dataset – 25 AI‑flavored inputs covering 25 distinct registers (tech blog, academic abstract, business email, LinkedIn post, README intro, etc.).
• Internal scorer – the repo’s own ai-check (rule‑based stylometry). Mean score = 5.24 / 27 (lower is more human); 6 Human, 19 Likely Human, 0 Uncertain or worse.
• External scorer – Binoculars (cross‑perplexity detector, ICML 2024). Mean score = 0.9928 (higher = more human) with all 25 outputs above the human threshold (0.854).
• Conclusion – Across diverse registers, the rewritten outputs are consistently judged human‑like by two independent, perplexity‑based detectors.

Known limits (as the authors state)

• The rule‑based approach cannot defeat learned classifiers such as GPTZero, Grammarly, or Pangram; those models look at the “RLHF fingerprint” baked into the original LLM weights, which surface rewriting alone cannot erase.
• The skill is static – it does not adapt or learn from new data. Its effectiveness depends on the nine levers staying relevant as detection research evolves.
• For high‑stakes scenarios where a learned‑classifier score matters, the README recommends complementary techniques (cross‑model paraphrase chains, base‑model rewriting, manual edits).

When you might want it

• You need a free, offline way to make AI‑generated drafts sound more personal before publishing or sharing.
• You want a quick, reproducible audit of a passage for AI‑generation signals without sending data to a third‑party service.
• You are already working inside a Claude Code, Codex CLI, or similar LLM‑agent environment and prefer a plug‑and‑play skill rather than an external API.

---

TL;DR – humanize provides two zero‑dependency, rule‑based skills that can be dropped into any LLM‑agent to (1) rewrite AI text so it reads like a human and (2) produce a detailed AI‑detection report. Installation is a single script or marketplace command, usage is a simple slash command, and the authors back the tool with a small benchmark showing human‑like scores on two perplexity‑based detectors. The approach is transparent and open‑source, but it does not beat modern learned AI‑detectors.