# sepia

An Agent Skill that rewrites AI-generated fiction and professional prose to remove the structural and stylistic patterns that reveal it as machine-written.

sepia is a portable Agent Skill that edits AI-generated writing to remove the patterns that make it identifiable as machine-written, rather than only swapping word choices. For fiction, it repairs narrative architecture first, such as how themes are explained and how revelations are paced, based on published research showing that a classifier can detect AI-written fiction from structural features alone even after human editors have reworked the surface style. Professional documents such as release notes, postmortems, and technical articles instead get rules matched to their specific venue.

It offers four operations: writing new text, reviewing existing text without changing it, a minimal-edit refactor, and a full rewrite, and applies a thirty-feature diagnosis rubric with per-model stylistic fingerprints when the writing or executing model is known. It installs through a Skills CLI that supports more than 77 agents, and ships native plugin packaging for Claude Code, Codex, Grok Build, Antigravity, and QwenPaw.

It is written in Python.

Website: https://github.com/Nanako0129/sepia