한국어 | 中文 | 日本語 | English

patina

Make AI text sound like a human wrote it.

A Claude Code skill and standalone CLI that detects and rewrites AI writing patterns in Korean, English, Chinese, and Japanese. Pattern-based, auditable, with deterministic scoring — not a black-box LLM paraphraser.

Demo

Before (AI-sounding):

Coffee has emerged as a pivotal cultural phenomenon that has fundamentally transformed social interactions across the globe. This beloved beverage serves as a catalyst for community building, fosters meaningful connections, and facilitates cross-cultural dialogue. From the bustling cafés of Paris to the serene tea houses repurposed for coffee in Tokyo, this remarkable journey showcases the innovative spirit of human culinary exploration.

After (/patina --lang en — same claims, AI packaging removed):

Coffee has quietly changed how people meet. Sit across from someone long enough, and something like a real connection tends to form — even between people from very different cultures. It happens in Paris cafés and in Tokyo tea houses that used to serve matcha. Somehow, one roasted bean grew into a social ritual shared across the world.

MPS = 100 · cultural transformation ✓ · community building ✓ · meaningful connections ✓ · cross-cultural dialogue ✓ · Paris cafés ✓ · Tokyo tea houses ✓ · culinary exploration ✓

At a Glance

Table of Contents

• Quick Start
• Modes & Flags
• MAX Mode
• Score & Ouroboros
• Authentication
• How It Works
• Calibration
• Patterns
• Configuration
• Profiles
• Custom Patterns
• Project Structure
• Adding a New Language
• References
• Version History

Quick Start

As a Claude Code skill

One-line install:

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

Then in Claude Code:

/patina --lang en

[paste your text here]

Manual install →

As a standalone CLI

Requires Node.js ≥ 18.

git clone https://github.com/devswha/patina.git
cd patina && npm install && npm link
patina --lang en input.txt

# Common usage
patina --lang en --profile blog input.txt
patina --lang ko --score input.txt
patina --lang en --ouroboros input.txt
patina --batch docs/*.md --suffix .humanized

🆓 No API key required if you have the codex CLI logged in. See Authentication for the full backend list.

Manual install

mkdir -p ~/.claude/skills
git clone https://github.com/devswha/patina.git ~/.claude/skills/patina
ln -snf ~/.claude/skills/patina/patina-max ~/.claude/skills/patina-max  # MAX mode skill

If you already cloned via the standalone CLI route, run npm link from that checkout instead of cloning twice.

Modes & Flags

patina --lang <ko|en|zh|ja> [mode] [--profile <name>] [batch options] input.txt

Combine freely: patina --lang en --audit --profile blog. See patina --help for all options.

MAX Mode (Multi-Model)

Run the same text through Claude, Codex, and Gemini independently. Each model humanizes, results are scored for AI-likeness + MPS, and the lowest-scoring (most human) result that passes MPS ≥ 70 wins.

/patina-max

[paste your text here]

Each MAX run uses an isolated temp directory, waits only for selected models, and marks timeouts as failures (no infinite waits).

Standalone CLI MAX: patina --models gpt-4o,gpt-4o-mini input.txt calls models via the same --base-url. To mix providers, point --base-url at OpenRouter or another multi-provider gateway. The Claude Code /patina-max skill dispatches via local CLIs — no API key needed.

Score & Ouroboros

Score mode

Check how AI-like your text is without rewriting:

patina --score input.txt

| Category      | Weight | Detected | Raw  | Weighted |
|---------------|--------|----------|------|----------|
| content       | 0.20   | 3/6      | 33.3 | 6.7      |
| language      | 0.20   | 1/6      | 11.1 | 2.2      |
| style         | 0.20   | 2/6      | 27.8 | 5.6      |
| communication | 0.15   | 0/3      | 0.0  | 0.0      |
| filler        | 0.10   | 1/3      | 11.1 | 1.1      |
| structure     | 0.15   | 1/4      | 25.0 | 3.8      |
| Overall       |        |          |      | 19.3 (±10) |

When combined with rewriting, also reports:

Ouroboros mode

Iterate the rewrite until score converges:

patina --ouroboros input.txt

| Iter | Before | After | Improvement | Reason     |
|------|--------|-------|-------------|------------|
| 0    | —      | 78    | —           | Initial    |
| 1    | 78     | 45    | +33         |            |
| 2    | 45     | 28    | +17         | Target met |

Termination conditions (whichever comes first):

• Target met (score ≤ 30, configurable)
• Plateau (improvement < 10 between iterations)
• Regression (score increased — rolls back)
• Max iterations (default 3)
• Fidelity / MPS floor breach (rolls back)

Configure in .patina.yaml:

ouroboros:
  target-score: 30
  max-iterations: 3
  plateau-threshold: 10
  fidelity-floor: 70
  mps-floor: 70

--ouroboros cannot be combined with --diff, --audit, or --score.

Authentication

patina auth status         # backend availability + auth state
patina auth login          # per-backend login instructions
patina --list-providers    # preset providers + key status

If PATINA_API_KEY is unset and codex is logged in, patina auto-falls back to codex-cli.

codex-cli v1 supports single-mode rewrites only. --audit, --score, --diff, --ouroboros, and --models/MAX still go through the HTTP backend.

Default environment variables:

PATINA_API_KEY=...                            # required for HTTP backend
PATINA_API_BASE=https://api.openai.com/v1     # or proxy
PATINA_MODEL=gpt-4o                           # default model

How It Works

Input text
  │
  ▼
[Step 4.5]   Semantic anchor extraction
             (key claims, polarity, causation, numbers)
  │
  ▼
[Step 4.6]   Stylometric pre-pass
             (burstiness CV + MATTR)
  │
  ▼
[Step 4.7]   AI-lexicon overlap
             (flat dictionary: ~108 EN / 102 KO entries)
  │
  ▼
[Phase 1]    Structure scan
             (paragraph-level: repetition, passive voice)
  │
  ▼
[Step 5a-v]  Anchor verification
  │
  ▼
[Phase 2]    Sentence rewrite
             (word-level: AI vocab, filler, hedging)
  │
  ▼
[Step 5b-v]  Anchor verification
  │
  ▼
[Phase 3]    Self-audit
             (polarity scan, regression, final MPS)
  │
  ▼
Natural-sounding text (meaning verified)

Pattern packs are auto-discovered by language prefix ({lang}-*.md). Semantic anchors are extracted before rewriting and verified after each phase — if meaning is corrupted, the change is retried or rolled back.

Calibration

Reproducible against .omc/research/v3_7_lexicon_eval.py over a 400-paragraph corpus (HC3 + Wikipedia + NamuWiki + paired ko/AI):

Acceptance gates: AI catch ≥ 75% · max FP ≤ 25% · NamuWiki regression ≤ +5pp. All met.

Stylometric and lexicon signals are advisory for the LLM, not sole-decision gates. The 25% Wikipedia FP is intrinsic to encyclopedic prose, not a bug we can tune away. Documented in core/stylometry.md §13, §16.

Patterns

All four languages share the same 6-category structure. Most patterns are universal; a few slots have language-specific variants. Patterns #30 (rhetorical question openers) and #31 (conclusion signal words) exist in all four languages. Pattern #32 (comparative adverb overuse — KO 보다, JA より) is KO/JA-specific.

Universal categories

Language-specific slots

Universal extensions (v3.4.0+)

Configuration

# .patina.default.yaml
version: "3.8.0"
language: ko              # ko | en | zh | ja
profile: default
output: rewrite           # rewrite | diff | audit | score
skip-patterns: []         # e.g., [ko-filler] to skip a pack
blocklist: []             # extra words to flag
allowlist: []             # words to never flag
max-models: [claude, gemini]
dispatch: omc             # omc | direct

Pattern packs are auto-discovered by language prefix — no need to list them manually.

Profiles

patina --profile blog text...

Custom Patterns

Drop a .md file into custom/patterns/ — auto-loaded:

---
pack: my-patterns
language: ko
name: My Custom Patterns
version: 1.0.0
patterns: 1
---

### 1. Pattern Name
**Problem:** What AI does wrong
**Before:** > AI-sounding example
**After:** > Natural-sounding fix

Project Structure

patina/
├── SKILL.md                  # /patina entrypoint
├── SKILL-MAX.md              # MAX mode reference
├── patina-max/               # /patina-max skill (installable)
│   └── SKILL.md
├── .patina.default.yaml      # Configuration
├── core/
│   ├── voice.md              # Voice & personality guidelines
│   ├── scoring.md            # Scoring algorithm reference
│   └── stylometry.md         # Stylometric algorithm reference
├── lexicon/
│   ├── ai-en.md              # English AI lexicon (108 entries)
│   └── ai-ko.md              # Korean AI lexicon (102 entries)
├── patterns/
│   ├── ko-*.md               # Korean (6 packs, 32 patterns)
│   ├── en-*.md               # English (6 packs, 31 patterns)
│   ├── zh-*.md               # Chinese (6 packs, 31 patterns)
│   └── ja-*.md               # Japanese (6 packs, 32 patterns)
├── profiles/                 # Tone presets
├── examples/                 # Before/after test cases
└── custom/                   # User extensions (gitignored)

Inspired by oh-my-zsh's plugin architecture: patterns are plugins, profiles are themes.

Adding a New Language

1. Create patterns/{lang}-content.md, {lang}-language.md, etc.
2. Set language: {lang} in each file's frontmatter.
3. Use /patina --lang {lang} — auto-discovered, no config changes.

References

• Wikipedia: Signs of AI writing — primary source
• WikiProject AI Cleanup — community effort
• blader/humanizer — original English version

Contributing

See CONTRIBUTING.md. Pattern submissions and staleness reports ("this signal isn't AI anymore") are the highest-value contributions — AI writing patterns evolve as models get fine-tuned.

Open an issue →

Version History

3.8.0 — Data-driven Korean lexicon mining

v3.7.0's Korean lexicon was author-curated and contributed only +1pp on AI catch (vs +10pp on English). v3.8.0 mines the corpus for high-signal Korean phrases via differential frequency against NamuWiki human prose, surfacing 12 register markers AI text uses heavily but humans rarely.

Mining rule (.omc/research/v3_8_ko_lexicon_mine.py):

• 어절 doc-frequency: AI count ≥ 4 AND ratio AI / (human + 1) ≥ 4.0
• Reject domain artifacts (proper nouns, year-tokens)
• Keep only register markers (passive evaluation, encyclopedic verbs, quantifier scaffolding)

Added entries:

• Strict (8): 평가된다, 꼽힌다, 가리킨다, 사례로, 다수의, 알려져, 일컬어진다, 평가받다
• Phrases (4): 가운데 하나로, 자리 잡았다, 알려져 있다, ~의 사례로

Result on 500-paragraph corpus: ko/AI catch 83% → 91% (+8pp). NamuWiki human FP held at 13% — zero regression, clean Pareto win.

3.7.0 — AI-lexicon overlap signal

A flat dictionary (lexicon/ai-en.md 108 entries, lexicon/ai-ko.md 90 entries) flags AI-favored phrases the 28-pattern catalog does not enumerate. Densities computed per 1000 tokens; the 4.6 hot rule extends to a 3-signal OR (burstiness OR MATTR OR lexicon_density > 2.0).

Calibrated against 400 paragraphs: AI catch 66% → 76%, HC3 human FP 12%→19%, Wikipedia FP 23%→25% boundary, NamuWiki FP 11%→13% (within +5pp guardrail). All acceptance gates met — first Pareto improvement over the v3.5.1 wall.

Drop list (post-eval): intersection, principles, mindset, iterative, responsible, methodologies, redefine, accessible, equitable, one of the most, in conjunction with, the power of — fired more on academic prose than on AI text.

Skipped v3.6 (n-gram dropped, §15 negative finding).

3.5.1 — Stylometric calibration patch

Raised stylometry.burstiness.bands.low from 0.25 to 0.30 after external validation against 300 paragraphs. v3.5.0 only caught 57% of real AI text; v3.5.1 catches 66% with HC3 human FP 12% and Wikipedia FP 23%.

Sweep showed no threshold combo satisfies both AI ≥70% and max FP ≤20% — Wikipedia's encyclopedic register naturally has uniform sentence length. MATTR threshold unchanged (0.55). v3.5.x is an advisory marker for the LLM, not a sole-decision gate.

License

MIT