patina is a multilingual AI writing humanizer that supports Korean (KO), English (EN), Chinese (ZH), and Japanese (JA). Its core positioning is "strip away the AI packaging, preserve the original meaning": it identifies expressions with an AI tone and rewrites them, but does not change the original claims, numbers, polarity, or causal relationships. The project explicitly states that it is intended for scenarios where "AI-assisted drafting is allowed," not for evading detectors.
Main capabilities and features
- Deterministic, pattern-based: every edit is driven by 189 named patterns, which the author describes as "auditable, not a black box"; use --diff to see the specific changes and their reasons.
- Semantic fidelity verification: on the web side, every rewrite goes through an MPS/fidelity lower-bound check, rejecting semantic drift; the CLI's --verify and /patina --strict add further checks.
- Three independent axes: Document Type (determines genre), Persona (determines tone), and Register (determines mode of expression). The three are mutually independent; omitting any axis preserves the corresponding feature of the original text; explicitly assigning a value does not automatically fill in the omitted axes.
- Multiple forms: agent skill (supports Claude Code, Codex, Cursor, OpenCode), Node CLI, and a browser playground.
- The project's own description is honest about limitations: scores are only an editorial reference signal and cannot serve as proof of authorship; it also publicly shares preregistered research, showing both cases where rewriting failed and cases where it was effective.
How to use
- Browser: no installation required, open patina.vibetip.help and paste the text. Rewriting is performed server-side with MPS/fidelity gates; in API mode, the user's own key is forwarded per request (not stored, not logged).
- Agent skill: can be pasted into Claude Code, Codex CLI, Cursor, or any agent; if you already use the skills.sh CLI, you can run npx skills add devswha/patina (which installs the full checkout, for use by 75+ agents).
- CLI: requires Node >= 18. If you are already logged in to codex, claude, or gemini CLI, you canadd --backend codex-cli to avoid needing an API key. See INSTALLATION.md for full installation options. patina --help prints all parameters. The GitHub Actions CI wrapper is devswha/patina-action, and there are also pre-commit, Docker, and release guides.
- The hosted API offers a Pro version (hosted service).
Project configuration is located in .patina.yaml.
Applicable scenarios and limitations
- Applicable: AI-assisted writing polishing in the four languages Korean/English/Chinese/Japanese, removing obvious AI tone while requiring facts, numbers, and logical relationships to be preserved; situations that require an auditable, explainable rewriting process.
- Limitations and notes: the project emphasizes that scores are an editorial signal with false positives and false negatives, and are not proof of authorship (see its Ethics section for details); the editor clients (VS Code, Obsidian, Gmail preview client) are no longer maintained; npm publishing is a separate step requiring explicit authorization, see release channels for details; the version number is 9.1.0. The license is MIT.
Other materials The project was inspired by oh-my-zsh, Wikipedia's "Signs of AI writing," and blader/humanizer. The repository also provides a Cookbook (common recipes), CLI contract (flags, gates, exit codes), Before/After gallery (Korean), pattern catalog, architecture, configuration and authentication, benchmarks, research, FAQ (Korean), contribution guide (Korean), and changelog. Submitting patterns requires using CONTRIBUTING.md: first submit a pattern proposal in .github/ISSUE_TEMPLATE/pattern_proposal.yml, then add it under patterns/{lang}-*.md according to the template.
