sepia Review 2026 — De-AI Writing Skill That Repairs Narrative Architecture, Not Just Word Choice
✅ Pros
- • Evidence-driven instead of folklore: the StoryScope study (arXiv:2604.03136, 61,608 stories from humans plus five frontier LLMs) showed a classifier using narrative-structure features alone detects AI fiction at 93.2% macro-F1 — and that editing surface style barely moves detection (95.5% → 93.9%). sepia attacks the architecture, not the adjectives
- • It names the actual tells: themes explained by the narrator, single-track causally-tidy plots, emotions rendered only as bodily sensation, no real-world references, no reader, linear time, endings resolved by protagonist growth and acceptance
- • A real protocol, not a prompt: three passes (narrative architecture → discourse flow → surface style), a 30-feature diagnosis rubric, and per-model fingerprint corrections for Claude, GPT, Gemini, DeepSeek, and Kimi
- • Professional prose gets its own rules: release notes (user impact first, artifacts per claim), PR/issue replies (answer first, cite file:line, no reflex praise, length ∝ stakes), postmortems (blameless toward people, merciless toward mechanisms), tickets (title = outcome, testable acceptance criteria), technical articles (one real dead end, one committed opinion, numbers with conditions)
- • Calibration principle is sound: 'calibrate to the human distribution, don't invert the AI one' — the skill selects 3–5 moves per story and leaves slack, explicitly avoiding a new uniform 'sepia' fingerprint from over-application
- • Truly portable: one canonical SKILL.md following the Agent Skills standard, packaged as plugins for Claude Code, Codex, Grok Build, and Antigravity, with four operations (write, review, refactor, recreate) and a research/ folder digesting 12+ cited studies
⚠️ Cons
- • Value depends entirely on the host agent: sepia is a skill, so output quality tracks the Claude Code / Codex / Grok Build / Antigravity model doing the rewriting — it can't fix a weak writer underneath
- • The 30-feature rubric and three-pass protocol take real judgment to apply; a rushed pass that applies every rule produces exactly the uniform-fingerprint failure mode the project warns about
- • No detector-evasion guarantee — and it's not the goal: the project is about narrative architecture and human-distribution calibration, and using any humanizer to defeat AI detectors raises academic-integrity questions the README doesn't engage with
- • Per-model fingerprint corrections are heuristic mappings, not per-model measured evaluations — reasonable starting points, but unvalidated for any specific model version
- • Heavy evidence base: the research/ digests are genuinely useful but dense, and non-English documentation currently exists only in Traditional Chinese (README.zh-TW)
- • Very young project (created August 28, 2026, 918 stars in three days) with a single canonical maintainer's voice shaping all four operations
Writers and technical communicators who produce long-form fiction or professional prose inside AI coding agents and want a research-grounded, multi-pass revision protocol — especially those writing release notes, postmortems, PR replies, or technical articles that must not read like chatbot output
Free, MIT-licensed; runs as an Agent Skill inside Claude Code, Codex, Grok Build, or Antigravity (your existing agent subscription/API costs apply)
Why “Another Humanizer” Is the Wrong Frame
Every popular AI-humanizer edits word choice and syntax: swap the clichés, break up the sentence templates, vary the vocabulary, adjust the register. sepia (repo Nanako0129/sepia, MIT, created August 28, 2026 — 918 stars in three days) makes a specific, research-backed argument that this is working on the wrong layer.
The load-bearing evidence is StoryScope (Russell et al., arXiv:2604.03136): a corpus of 61,608 stories written by humans and five frontier LLMs, used to train a classifier that distinguishes them. The result that drives sepia’s design: a classifier using narrative-structure features alone detects AI fiction at 93.2% macro-F1 — and editing the surface style away barely moves it (95.5% → 93.9%). The tells that survive surface edits are architectural: themes explained by the narrator, single-track causally-tidy plots, emotions rendered only as bodily sensation, no real-world references, no reader, linear time, endings resolved by protagonist growth and acceptance.
“De-AI writing at the layer that actually gives AI away.” — sepia README
So sepia’s answer to “another humanizer?” is: the others fix the words; this fixes the structure that makes the words a giveaway in the first place.
The Protocol: Three Passes, Four Operations, One Rubric
sepia is a portable Agent Skill (per the Agent Skills specification) — one canonical SKILL.md, no per-platform forks — installable as a plugin in Claude Code, Codex, Grok Build, and Antigravity, with four operation entries: write (create new prose), review (diagnose only, no edits), refactor (minimal in-place edits), and recreate (full rewrite from source facts and intent).
The writing/revision protocol is a three-pass stack, each pass targeting a different layer:
| Pass | Layer | Example moves |
|---|---|---|
| 1 | Narrative architecture (fiction) | stop explaining the theme, loosen the causal chain, back-load revelations, mix emotion modes, sparse character networks, name real things |
| 2 | Discourse flow | de-template the paragraph-question sequence, fix the mid-story sag, vary rhythm and positions |
| 3 | Surface style | the classic layer: clichés, syntax templates, vocabulary, register |
Sitting on top is a 30-feature diagnosis rubric and per-model fingerprint corrections for Claude, GPT, Gemini, DeepSeek, and Kimi — the idea being that each model family has its own detectable defaults, so the fix should be model-aware.
Professional Prose Fails Differently — So It Gets Its Own Rules
sepia’s most distinctive feature is that it doesn’t treat all writing the same. The research base (12+ digested studies, including LAMP, Measuring AI Slop, NarraBench, and Echoes in AI) points at different tells for professional documents: filler that carries no information, hedging where a judgment was needed, chatbot leftovers, register that ignores the venue, formatting that looks stamped out. Each document type gets a thin rule file on top of one shared checklist:
- Release notes / announcements: user impact first, artifacts per claim, no marketing inflation
- PR / issue replies: answer first, cite
file:line, no reflex praise, length proportional to stakes - Postmortems: blameless toward people, merciless toward mechanisms — timestamps, dead ends, owned action items
- Tickets / work orders: title = outcome, testable acceptance criteria, link don’t repeat
- Technical articles: open at the problem, one real dead end, one committed opinion, numbers with conditions
That last row is worth reading twice — “one real dead end, one committed opinion, numbers with conditions” is a better brief for technical writing than most style guides give.
The Calibration Principle: Don’t Invert the AI Distribution
The governing principle, stated explicitly, is the project’s most sophisticated idea: calibrate to the human distribution, don’t invert the AI one.
Humans sit at moderate values on most features. A story with every rule applied is a new fingerprint — an AI that has been “de-AI’d” with maximum effort is statistically just as detectable, because the output now over-corrects. So sepia deliberately selects 3–5 moves per story and leaves slack, applying only the highest-value corrections and leaving some tells (or near-tells) in place, the way a human writer actually writes. It also means the tool’s job is not detector evasion — it’s producing prose that sits in the human distribution because it was built with human structural habits.
Practical Realities
Installation is a one-liner per platform: npx skills add Nanako0129/sepia -g (the Skills CLI supports 77+ agents), or native plugin commands (claude plugin install sepia@sepia, codex plugin remove sepia@sepia, etc.). Project-scoped installs commit the skill into .claude/skills/sepia or .agents/skills/sepia.
The honest limitations: as a skill, sepia’s output quality is capped by the host model doing the rewriting — it amplifies a good writer and inherits a weak one. The three-pass protocol demands judgment; a mechanical pass that applies every rule produces exactly the uniform fingerprint the project warns against. The per-model fingerprints are heuristic mappings, not measured per-model evaluations. And while the research base is a genuine strength, its density is a barrier for casual users.
Compared to alternatives: classic humanizers (Undetectable AI, GPTZero-style rewriters) work on surface style and target detectors — sepia explicitly does not promise evasion and works on structure; writing-style skills like the Refactoring UI skill for design target a different domain entirely. For writers inside coding agents, sepia occupies a genuinely empty niche: a research-grounded, multi-pass revision protocol for narrative and professional prose.
Verdict
sepia is the rare AI-tooling project that earns its claims with citations. The StoryScope numbers are specific and reproducible, the three-pass protocol is actionable, the per-domain rule files are sharp, and the calibration principle shows actual thinking about the failure mode of over-correction. It’s young, it’s opinionated, and it lives or dies by the host agent’s writing ability — but for anyone producing long-form or professional prose inside Claude Code, Codex, Grok Build, or Antigravity, it’s the most substantive De-AI writing tool to ship this year.
Screenshots captured from the official GitHub repository on August 31, 2026. Star counts and metrics reflect the repository state at review time.