sepia — Making AI Writing Less Detectable

What it is

sepia is a set of instructions (an "Agent Skill") that AI coding agents like Claude Code, Codex, or Grok Build can load to write text that doesn't sound like AI wrote it. It targets fiction and professional documents (release notes, PR replies, postmortems, tickets, technical articles).

The core idea

Most "humanizer" tools just swap out words and sentence patterns. But research cited in the README (the StoryScope study, arXiv:2604.03136) found that AI fiction is detectable mainly by narrative structure, not surface style — things like the narrator explaining the theme, overly tidy cause-and-effect plots, emotions described only as physical sensations, no real-world references, and endings resolved by the protagonist's personal growth. Even after human editors rewrote the surface style, detection barely dropped (95.5% → 93.9%).

So sepia works in three passes:

1. Narrative architecture — fix structural tells (stop explaining the theme, loosen the causal chain, mix emotion modes, name real things)
2. Discourse flow — fix paragraph patterns, mid-story sag, rhythm
3. Surface style — the classic layer: clichés, vocabulary, register

A key principle: it doesn't apply every rule to every piece. It selects 3–5 targeted moves per story, because a story with all rules applied becomes its own detectable fingerprint.

How you use it

Five operations, each a command like /sepia-write or $sepia-review:

Operation	What it does
write	Create new prose
review	Diagnose AI tells without editing
refactor	Make minimal in-place edits
recreate	Full rewrite from source facts and intent
hemingway	Write/refactor fiction with a built-in Hemingway voice

Professional documents

Different document types get different rule files: release notes should lead with user impact, PR replies should answer first and cite file:line, postmortems should be blameless toward people but merciless toward mechanisms, tickets need testable acceptance criteria.

Voice skills (experimental)

Since v0.4.0, you can stack a voice or style skill on top — a minimalism method, a brand voice, a persona. Built-in profiles include Hemingway, Taiwan long-form journalism, and a persona template. The interface has strict precedence rules: sepia's architecture decisions always come first, and voice moves are applied selectively (3–5 per piece).

Chinese text support

sepia calibrates its style pass for Chinese using four evidence layers: the HC3 corpus, a private corpus of ~2,000 Taiwanese journalism articles, a machine-side contrast group, and Taiwan writing conventions. The one syntactic measure it trusts across languages: within-passage variance in sentence length is consistently higher in human text. Mean sentence length, punctuation counts, and paragraph length were all discarded as unreliable signals.

Installation

• Any agent: npx skills add Nanako0129/sepia -g (works with 77+ agents via the Skills CLI)
• Native plugins: Claude Code, Codex, Grok Build, Antigravity, and QwenPaw each have dedicated install commands

Research base

The research/ directory digests ~15 academic papers on AI text detection, including StoryScope, LAMP, and several PNAS/ACL studies. The README is careful to note which claims are measured vs. anecdotal — e.g., the voice-skills interface is grounded in "one blind review experiment... a worked example, not measured evidence."

Bottom line

sepia is a well-researched, multi-platform tool that targets the structural tells of AI writing rather than just surface style, with a careful "calibrate to the human distribution, don't invert the AI one" philosophy.