De-AI Your Writing With sepia's Three-Pass Protocol

Ask a model to "make this sound less like AI" and you get word swaps — while the tells that actually give the text away stay untouched. sepia, a 3k-star agent skill, exists because of what the detection research shows: in the StoryScope study it cites (arXiv:2604.03136), a classifier using narrative-structure features alone flagged AI fiction at 93.2% macro-F1, and when human editors rewrote the surface style, detection only dropped from 95.5% to 93.9%. The tells that survive editing are architectural.

Why This Skill Matters

sepia works in three passes, each on a different layer:

1. Narrative architecture (fiction) — stop explaining the theme, loosen the causal chain, back-load revelations, mix emotion modes, name real things.
2. Discourse flow — de-template the paragraph-question sequence, fix the mid-story sag, vary rhythm.
3. Surface style — the classic layer other tools stop at: clichés, syntax templates, vocabulary.

Professional prose gets its own venue-matched rule files on top of one shared checklist — release notes lead with user impact, PR replies answer first and cite file:line, postmortems stay blameless toward people and merciless toward mechanisms, tickets get testable acceptance criteria, technical articles open at the problem, long-form journalism splits lead and body into two registers.

One governing principle keeps it from over-correcting: calibrate to the human distribution, don't invert the AI one. The skill selects 3–5 moves per piece and leaves slack, because a story with every rule applied is a new fingerprint.

Installation

The Skills CLI route covers 77+ agents — install once at user scope:

npx skills add Nanako0129/sepia -g

For Claude Code, the README documents the native plugin route:

claude plugin marketplace add Nanako0129/sepia
claude plugin install sepia@sepia --scope user

Native plugin packaging also exists for Codex, Grok Build, Antigravity, and QwenPaw; the README's honesty note is worth heeding — "verified" there means the install completes and the entries appear, not that every behavior was exercised per platform. One constraint: install the complete plugin package. The README documents that standalone wrapper installation is unsupported because the wrappers depend on their sibling canonical skill.

Real Workflow: Diagnose, Then Repair a Release Note

Write your draft first — sepia's review operation diagnoses without editing. With the skill installed:

用 /sepia-review 诊断这份发布说明：按它的评分表列出叙事架构、篇章推进和表层文风三个层面的 AI 痕迹，只诊断，不要改文字。

Review reports where the tells cluster — for the README's own worked example (a maintainer's Threads post), they sat at the opening and the close: an opening about how the change came to be, a maxim ending, a few absolutes. Then switch operations:

• /sepia-refactor makes minimal in-place edits — the maintainer's post kept its middle untouched.
• /sepia-recreate rewrites from the source facts and intent — a full rewrite.
• /sepia-write creates new prose with the protocol applied from the start.
• /sepia-hemingway applies the built-in Hemingway voice — iceberg omission for fiction, the Kansas City Star rules for professional prose.

For Chinese text, the style pass loads a dedicated calibration (references/languages/zh.md) built on four evidence layers: the HC3 corpus, about two thousand articles from one Taiwanese publication spanning roughly ten years, a 119-piece machine-written contrast group, and Taiwan writing conventions.

Tips

• Fiction and professional docs fail differently — fiction needs the architecture pass; release notes need the venue rules. Let the router pick, or call the operation directly.
• Voice stacking is opt-in and experimental: tell sepia a voice skill is in play and it loads that profile's moves selectively — 3–5 signature moves, not a full rewrite in the voice.
• Review keeps full strength on uniformity findings even with a voice loaded — the README puts it sharply: a voice does not excuse a metronome.
• The research/ directory digests the primary studies behind every rule, with arXiv links.

When Not to Use This

sepia is a repair protocol, not a detector-evasion guarantee — the README cites detection studies as motivation, and it promises calibrated writing, not undetectable output. If your draft already reads well, running review costs a pass and may return little; the skill's own worked example only patched the opening and close. And if you want a persistent persona voice across documents, expect to write the persona profile yourself — the built-ins are Hemingway, a Taiwan journalism voice, and one companion persona; the README ships a template and a checker for author-created profiles.

See the leaderboard for more skills.