patina

Strip the AI packaging. Keep the meaning.

한국어 · 中文 · 日本語 · English

Try it in the browser — no install

Paste AI-sounding text → patina rewrites it in place, keeps the facts (the "30 templates" number survives), and drops the deterministic AI signal 100 → 0 (MPS 100 / Fidelity 75).

The words are not the tell. The architecture is. We ran a pre-registered study with independent cross-family LLM judges: a rewrite can strip the AI vocabulary, yet judges still recognize AI documents by their shape — uniform paragraphs, checklist-complete coverage, the tidy problem→lesson arc. Word-level cleanup moved judged AI-likeness a lot on English documents (−23 points) and barely at all on long Korean ones. That result decides patina's design: detect both layers, and publish the losses next to the wins (the study).

patina is a deterministic, pattern-based humanizer for Korean, English, Chinese, and Japanese. It finds AI-sounding phrasing and rewrites it without changing the claim, numbers, polarity, or causation. It is not a black-box paraphraser, authorship detector, or detector-bypass tool — it is built for allowed AI-assisted drafting where the author wants cleaner voice, an audit trail, and meaning-preservation checks.

More examples: Before/After Gallery (한국어) · CLI transcript.

Quick Start

Browser playground

Open patina.vibetip.help — paste KO / EN / ZH / JA text for a real rewrite gated by the MPS/fidelity floors, with the deterministic AI signal measured before → after. Rewrites and scoring run server-side; the free tier uses the service's own model key (rate-limited). API mode forwards your own key per request through the patina server to the provider you pick — never stored or logged (metrics are sanitized: no text, prompt, output, key, or IP). A hosted Pro tier unlocks higher limits — see pricing.

Agent skill

Let your coding agent install it — paste this into Claude Code, Codex CLI, Cursor, Gemini CLI, or any agent:

Install patina by following https://raw.githubusercontent.com/devswha/patina/main/INSTALLATION.md

The agent fetches INSTALLATION.md (written for AI agents) and runs the right install path for your host, then verifies it. Or do it yourself:

Claude Code — plugin marketplace (no clone, recommended):

/plugin marketplace add devswha/patina
/plugin install patina@patina

Claude Code · Codex CLI · Cursor · OpenCode — install script:

curl -fsSL https://raw.githubusercontent.com/devswha/patina/main/install.sh | bash

Then run the skill from Claude Code, Codex CLI, Cursor, or OpenCode:

/patina --lang en

[paste your text here]

Useful skill calls:

/patina --document-type email --register professional
/patina --document-type blog --persona pragmatic-founder

Standalone CLI

Requires Node.js >= 18.

npx patina-cli doctor
npx patina-cli --lang en input.txt

Use a logged-in local model CLI without an API key:

printf '%s\n' 'Coffee has emerged as a pivotal cultural phenomenon.' \
  | npx patina-cli --lang en --backend codex-cli

Supported local backends: codex-cli, claude-cli, gemini-cli, kimi-cli — patina passes the strongest documented default model per backend. See Authentication (한국어).

For large --batch runs, prefer an OpenAI-compatible HTTP backend; local CLI backends are agent runtimes, capped conservatively with --timeout-ms, --max-concurrency, --max-retries, and --max-failures for batch safety.

What You Get

184 patterns	37 rewrite-capable + 9 score-only viral-hook per language (46 each across KO/EN/ZH/JA) — see the full 184-pattern catalog in PATTERNS.md
Modes	rewrite · verify · audit · score · diff
Surfaces	agent skill · Node CLI · in-place preview · browser playground (rewrite + score)
Rewrite axes	--document-type controls genre/policy · --persona controls reusable voice · --register controls casual/professional delivery
Free usage	logged-in codex, claude, or gemini CLI can run rewrites without PATINA_API_KEY
Calibration	67.3% editing-hotspot catch [63.5–71.0%] across GPT-5.5 / Claude Sonnet 4.6 / Gemini 2.5 Pro (n=600, KO+EN); 16.0% false positives [11.6–21.7%] on KO+EN human controls (n=200)
License	MIT

Scores are editing signals with false positives and false negatives, not proof of authorship. See Ethics.

Three Independent Axes

Patina does not infer one axis from another. Omit Persona and Register to keep the source voice and register.

Axis	Controls	Does not control	CLI	Config	Playground
Document Type	Genre, purpose, structural conventions, pattern policy	Voice, casual/professional delivery, meaning floors	--document-type	document-type	Document
Persona	Reusable voice fingerprint: vocabulary, rhythm, explanation habits	Genre, pattern policy, register, meaning floors	--persona	persona	Voice
Register	casual or professional delivery	Genre, persona identity, pattern policy	--register	register	Register
Meaning preservation is the outer guard. If instructions appear to overlap,
field ownership resolves them: Document Type wins for document structure and
domain constraints, Persona for idiolect and rhythm, and Register for
casual/professional markers. An explicit value never fills an omitted axis.

Examples:

patina --document-type email --register professional note.md
patina --document-type blog --persona pragmatic-founder post.md
patina --document-type technical --persona technical-explainer --register casual guide.md

Common Commands

patina --lang <ko|en|zh|ja> [mode] [--document-type <name>] [--persona <name>] [--register <name>] input.txt

Command	Purpose
patina input.txt	rewrite with defaults
patina --audit input.txt	detect patterns only
patina --score input.txt	output a 0-100 score using the LLM judge plus deterministic signals
patina --score --offline input.txt	score with deterministic signals only; no backend or API key
patina --score --offline --exit-on 30 input.txt	offline CI gate with exit code 3 when overall > 30
patina --diff input.txt	show pattern-by-pattern changes
patina --preview page.html	render rewrites back onto a saved HTML page with toggles and inline diff
patina --verify input.txt	rewrite, then check MPS/fidelity floors with one retry
patina --document-type email --register professional input.txt	use email conventions with professional delivery
patina --persona pragmatic-founder input.txt	rewrite in a built-in voice persona
patina persona new my-voice --from-sample past.txt	author your own persona from a writing sample
patina persona list	list built-in + custom personas
patina persona show my-voice --json	print a persona's normalized config (never the docs body)
patina persona edit my-voice --name "New Name"	copy-on-edit a persona into custom/personas/
patina persona rm my-voice	remove a custom persona (built-ins are protected)
patina --format json --quiet input.txt	script-friendly output
patina --batch docs/*.md --outdir cleaned/	batch file processing

patina --help prints the full flag list. patina doctor --json checks Node, backend, tmux, and API-key readiness without making an LLM call.

Personas (voice)

A persona is a reusable voice — a built-in (patina persona list) or your own, authored without editing source:

patina persona new my-voice --from-sample past-posts.txt   # learn from your writing
patina persona new my-voice --describe "plain-spoken founder, casual"
patina --persona my-voice draft.md                          # then reuse it

patina persona show my-voice                                # inspect the normalized config (--json for machine output)
patina persona edit my-voice --name "Founder voice"         # copy-on-edit into custom/personas/ (built-ins stay intact)
patina persona rm my-voice                                  # remove a custom persona (--force to skip the confirm)

Works on ko/en/zh/ja and composes independently with --document-type and --register. Persona v2 changes reusable voice only; it cannot select a document type, set register, change pattern policy, or define verification/meaning floors. Authored Personas are validated on save, while global rewrite guards and --verify enforce meaning preservation independently.

CI

For GitHub Actions, the maintained wrapper is shorter than hand-rolled setup:

name: Patina prose score
on:
  pull_request:
    paths: ['**/*.md', '**/*.mdx']
permissions:
  contents: read
  pull-requests: read
  issues: write
jobs:
  patina:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v6
      - uses: devswha/patina-action@v1
        with:
          score-threshold: 30
          lang: auto
          comment: true

Other integrations: pre-commit, static sites, Docker, release workflow.

How It Works

Input
  -> semantic anchor extraction (claims, polarity, causation, numbers)
  -> stylometry + AI-lexicon scan
  -> pattern-guided rewrite
  -> self-audit and MPS/fidelity checks
  -> cleaned text

If meaning drifts, the change is retried or rolled back. Deterministic analysis lives in src/features/*; LLM-backed rewrite and score calls use the selected backend.

Configuration

# .patina.yaml
version: "7.0.0"
language: ko              # ko | en | zh | ja
document-type: default    # genre/purpose + pattern policy
persona:                  # optional reusable voice; omit to preserve source
register:                 # casual | professional; omit to preserve source

Project .patina.yaml overrides defaults. Pattern packs are auto-discovered by language prefix. Additive list keys (blocklist, allowlist, skip-patterns) merge; other arrays replace.

Documentation

Start here:

• Cookbook — common recipes and workflows
• CLI Contract — flags, formats, score gates, exit behavior
• Authentication — local CLI backends and API providers
• Patterns — full pattern catalog
• Subagents & strict flow — optional read-only detector/fidelity/naturalness subagents and the --strict multi-pass mode
• Benchmarks · latest report · 2026 rebaseline
• Measurement harness — index of every benchmark, calibration, and gate tool (incl. the signal-impact ablation harness)
• FAQ (한국어)
• Ethics
• Contributing (한국어)
• Changelog

Brand assets and usage rules live in Branding. Design notes live in DESIGN.md.

Acknowledgements

Inspired by oh-my-zsh's plugin architecture, Wikipedia's "Signs of AI writing", and blader/humanizer.

License

MIT. See LICENSE and NOTICE.