sepia

sepia is an open-source Agent Skill from Nanako0129 for de-AI writing. It works on fiction and on professional documents such as release notes, PR replies, postmortems, tickets, and technical articles. The latest release listed is v0.12.3, and the project is MIT licensed.

It ships as one canonical SKILL.md that any agent supporting the Agent Skills standard can load. The Skills CLI installs it, and Claude Code, Codex, Grok Build, Antigravity, and QwenPaw also get native plugin packaging.

What It Is

sepia is a writing and revision skill that runs inside an AI agent. It offers four operations: write (create new prose), review (diagnose only), refactor (minimal in-place edits), and recreate (full rewrite from source facts and intent). The plugin package adds a general router plus direct entries such as /sepia-write, /sepia-review, /sepia-refactor, /sepia-recreate, and /sepia-hemingway.

Three-Pass Protocol

The README says sepia is based on StoryScope (arXiv:2604.03136), which reports that narrative-structure features alone detect AI fiction at 93.2% macro-F1. The skill therefore works in three passes:

• Narrative architecture for fiction, such as not explaining the theme, loosening the causal chain, and naming real things.
• Discourse flow, such as de-templating paragraph sequences and varying rhythm.
• Surface style, covering clichés, syntax templates, vocabulary, and register.

A 30-feature diagnosis rubric and per-model fingerprints apply when the writing model is known. The governing principle is to calibrate to the human distribution rather than invert the AI one, so only 3-5 moves are selected per story.

Professional Writing Rules

Each document type gets a thin rule file on top of a shared checklist. Release notes lead with user impact. PR replies answer first and cite file:line. Postmortems are blameless toward people. Tickets use testable acceptance criteria. Technical articles open at the problem. Long-form journalism has its own rules.

Voice Skills and Chinese Calibration

An experimental, opt-in interface lets a voice or style skill stack on top of sepia. Built-in profiles include Hemingway, Taiwan long-form journalism, and a Nyaneko persona, and a persona template is checked by scripts/check_persona.py. Chinese text loads a language reference that calibrates the style pass using several evidence layers, including the HC3 corpus and Taiwan writing conventions.

Setup Path

Install for any supported agent with npx skills add Nanako0129/sepia -g. Native installers exist for Claude Code, Codex, Grok Build, Antigravity, and QwenPaw. A project-scope install is also possible by committing skills/sepia/ into a repo.