# code-humanizer <img width="1280" height="640" alt="social-preview" src="assets/social-preview.png" />

English | [中文](README.zh.md)

[](https://github.com/LeonardNJU/code-humanizer/stargazers) [](LICENSE) [](https://github.com/LeonardNJU/code-humanizer/commits) [](https://github.com/LeonardNJU/code-humanizer/issues)

**[humanizer](https://github.com/blader/humanizer), but for code.** An agent skill that removes signs of AI-generated code from a repository, and can guard new changes against the same patterns — the structural slop coding agents leave behind when they optimize for "tests pass" instead of "codebase stays healthy."

## Install

One command, picks your agents for you (Claude Code, Codex, Cursor, Cline, Gemini, Copilot, and more):

```bash npx skills add LeonardNJU/code-humanizer ```

Or drop it in by hand — the skill is a single `SKILL.md`, no build step, no dependencies:

```bash # Claude Code git clone https://github.com/LeonardNJU/code-humanizer ~/.claude/skills/code-humanizer ```

Any harness that reads `SKILL.md` agent skills works the same way.

## Use

``` > use code-humanizer to scan this PR > deslop pkg/report.py — you have my approval to fix > this repo was vibe-coded, humanize it (report first) > implement this change with code-humanizer guard > keep experiments/ exploratory, but guard changes under src/ ```

Default mode is **scan → report** (findings table with pattern #, severity 0–4, evidence, proposed fix, behavior risk). Fix mode runs on your approval, one pattern per commit, tests green after every step.

## Prevent slop while coding

**Guard mode** applies the same c