Sepia targets the structure of AI-written text
The open source skill rewrites narrative architecture rather than phrasing, citing a study that still flags AI prose at 93.2 percent macro-F1.
Illustrative image: overlapping manuscript pages on a desk, with a thin thread linking two pinned index cards.
Sepia is an MIT-licensed skill package for Claude Code, Codex, Grok Build, and Antigravity that reworks AI-sounding text at the level of narrative structure instead of sentence style.
At a glance
- Four operating modes: write, review (diagnosis only), refactor (minimal edits), recreate (full rewrite).
- Fiction protocol runs three passes: narrative architecture, discourse flow, surface style.
- A 30-feature diagnosis rubric, plus corrections for Claude, GPT, Gemini, DeepSeek, and Kimi.
- MIT license; 446 stars and 23 forks on GitHub at the time of retrieval.
- The repository grounds the approach in StoryScope (arXiv:2604.03136).
Sepia is an MIT-licensed skill package for Claude Code, Codex, Grok Build, and Antigravity that reworks AI-sounding text at the level of narrative structure instead of sentence style. Its stated reasoning is blunt: swapping words leaves untouched whatever detectors are actually measuring.
The premise
The repository points to a paper called StoryScope (arXiv:2604.03136). As the project reports it, AI fiction stays identifiable at 93.2 percent macro-F1 even after a stylistic revision pass. From that, Sepia concludes the giveaways are architectural, and that surface editing is the wrong lever.
That figure comes from the project's own description. The paper was not read for this article, so the setup and data behind the measurement are not verified here.
Four modes, three passes
Sepia offers four modes: write for new text, review for diagnosis with no edits, refactor for small targeted changes, and recreate for a full rewrite. Separating diagnosis from intervention is the practically useful part, because it leaves a step where a human decides what to change.
For narrative work, the skill runs three passes. The first handles architecture: less spelled-out theme, looser causal chains, revelations held back longer, a wider range of emotional expression, sparser character networks, references to the real world. The second addresses discourse flow, meaning templated paragraph sequences and sagging mid-story pacing. Only the third pass touches cliché, syntax patterns, vocabulary, and register.
Alongside that sit a 30-feature diagnosis rubric and per-model corrections for Claude, GPT, Gemini, DeepSeek, and Kimi — an assumption that different systems leave different recurring habits behind.
Rules for workplace prose
Beyond fiction, Sepia carries genre-bound rules for professional writing. Release notes lead with user impact. Pull request replies answer first and cite. Postmortems stay blameless about people and get rigorous about mechanisms. Tickets get outcome-focused titles. Technical articles start from the problem, name the dead ends, and commit to an opinion.
What is not established
Only one independent source was available for this article: the repository itself. No second newsroom describing or testing the project could be used. Everything about how Sepia works, and how well, is the project's own account rather than an independently confirmed result.
Three points in particular stay open. Whether the cited detection rate matches the paper the project leans on. Whether the 30 features actually make revised text harder to classify in practice. And whether the model-specific corrections survive past the model versions they were written against. Star and fork counts are a snapshot of the moment the page was retrieved.
The code is MIT licensed. Installation is offered per platform, through a single shell script, or as a project-scoped checkout, with user scope as the default.
FAQ
What does Sepia actually do?
It revises AI-generated text at the level of narrative structure, and can either diagnose only, apply minimal edits, or rewrite the piece completely.
Which tools does Sepia support?
The repository lists Claude Code, Codex, Grok Build, and Antigravity. By default it installs at user scope.
Is there proof Sepia defeats AI detection?
No. The project describes an approach and cites a study; no independent verification of its effectiveness is available here.
