Nanako0129/sepia

De-AI writing skill for any Agent Skills-compatible agent (77+ via the Skills CLI), with native plugins for Claude Code, Codex, Grok Build, and Antigravity. Narrative-architecture repair for fiction, venue-matched rules for professional prose. Based on StoryScope (arXiv:2604.03136)

Summary

sepia is an open-source Agent Skill that rewrites AI-generated prose to remove detectable AI patterns, with native plugins for Claude Code, Codex, Grok Build, Antigravity, and QwenPaw, plus a portable Skills CLI installer for 77+ agents. Unlike word-level humanizers, it works in three passes — narrative architecture, discourse flow, and surface style — grounded in StoryScope research showing that structural features alone detect AI fiction at 93.2% macro-F1, while surface edits in the LAMP condition only shifted detection from 95.5% to 93.9%. It also ships venue-matched rule files for professional writing: release notes, PR replies, postmortems, tickets, technical articles, and long-form journalism.

What it is useful for

• Revising LLM-generated fiction to fix the architectural tells the research flags: themes explained by the narrator, single-track causally-tidy plots, emotions rendered only as bodily sensation, missing real-world references, linear time, and endings resolved through protagonist growth and acceptance.
• Polishing professional AI-assisted writing where the failure modes differ. Each venue gets a thin rule file: release notes put user impact first and avoid marketing inflation, PR replies answer first and cite file:line, postmortems stay blameless toward people but merciless toward mechanisms, and tickets get outcomes in the title with testable acceptance criteria.
• Diagnosing text without changing it (/sepia-review), making minimal in-place edits (/sepia-refactor), or rewriting from source facts (/sepia-recreate).
• Layering an opt-in voice or persona on top — a built-in Hemingway profile or a Taiwan long-form journalism style — through an experimental voice-skills interface, which applies 3–5 signature moves per piece rather than every rule.

The README also warns that sepia is not aimed at maximising rule coverage: the governing principle is to "calibrate to the human distribution, don't invert the AI one," since a story with every rule applied becomes its own fingerprint.

How engineers can use it

The README installs the complete plugin package only; standalone wrapper install is unsupported. For most agents the one-liner is:

npx skills add Nanako0129/sepia -g

Platform-specific installers documented in the README:

• Claude Code: claude plugin marketplace add Nanako0129/sepia then claude plugin install sepia@sepia --scope user
• Codex: codex plugin marketplace add Nanako0129/sepia then codex plugin add sepia@sepia
• Grok Build: grok plugin install Nanako0129/sepia --trust
• Antigravity: agy plugin install https://github.com/Nanako0129/sepia
• QwenPaw: git clone then qwenpaw plugin install ./sepia/.qwenpaw-plugin

Invocation matches the standard form across hosts (/sepia-write, /sepia-review, /sepia-refactor, /sepia-recreate, /sepia-hemingway). A persona template at references/voices/PERSONA-TEMPLATE.md lets teams write down a house voice once; the validator at scripts/check_persona.py checks its structure.

Documented limitations worth knowing before adoption: "verified" means the install completes and entries appear, not that behaviour has been checked platform by platform, and runtime outside the five native plugins has not been exercised. Per-model fingerprints exist for Claude, GPT, and Gemini, while DeepSeek and Kimi are listed as consulted because no prompt guidance is published. The voice-skills interface is grounded in one blind review on a strict-minimalism specimen — a worked example, not measured evidence. The project is MIT-licensed.