Nanako0129/sepia

English | 繁體中文 | 简体中文

De-AI writing at the layer that actually gives AI away. Fiction gets its narrative architecture repaired before anyone touches word choice; professional documents (release notes, PR replies, postmortems, tickets, technical articles) each get rules matched to their venue.

A portable Agent Skill: any agent that speaks the standard can load it, and the Skills CLI, which supports 77+ agents, installs it with one command. Claude Code, Codex, Grok Build, Antigravity, and QwenPaw additionally get native plugin packaging. One canonical SKILL.md, no per-platform forks. Four operations: write, review (diagnose only), refactor (minimal edits), recreate (full rewrite).

• Why another humanizer
• Operation entries
• Experimental: composing with voice skills
• Sentence rhythm and Chinese calibration
• Install
• Uninstall
• Layout
• Star History
• Sources
• Support
• License

Every popular humanizer edits word choice and syntax. StoryScope (Russell et al., 2026: 61,608 stories, human + 5 frontier LLMs) showed that a classifier using narrative-structure features alone detects AI fiction at 93.2% macro-F1. In the same study's LAMP-edited condition, where human editors had rewritten the surface style, detection dropped only from 95.5% to 93.9%. The tells that survive 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.

sepia turns those measured gaps, together with the related studies digested in research/, into a three-pass writing and revision protocol for fiction:

A 30-feature diagnosis rubric and per-model fingerprints across two layers apply when the writing or executing model is known:

Notice: Vendors that publish no prompt guidance are recorded as consulted, not guessed.

Professional prose fails differently, and the structure-level finding holds there too: a 2026 replication of StoryScope on 2,250 company blog posts against 11,250 AI mirrors separated them at 98.0 macro-F1 from structural features alone, with the AI shape described as tidy and self-announcing (SLOPSHAPE-2026 in the ledger, arXiv:2609.15369; a preprint whose features are LLM-scored; it tested detection of original and model-self-reworded posts, never human editing). The studies digested in research/ point at filler that carries no information, hedging where a judgment was needed, chatbot leftovers, register that ignores the venue, and formatting that looks stamped out. Each document type gets a thin rule file on top of one shared checklist:

Governing principle: Calibrate to the human distribution, don't invert the AI one. Humans sit at moderate values; a story with every rule applied is a new fingerprint. The skill selects 3–5 moves per story and leaves slack.

The complete plugin package gives Claude Code, Codex, Grok Build, Antigravity, and QwenPaw a general router plus five direct entries. On QwenPaw the package registers no commands of its own; the host's /<skill-name> dispatch serves each installed skill:

The general /sepia (Claude Code, Grok Build, Antigravity, and QwenPaw) or $sepia (Codex) router remains available. What was verified on each platform is stated under Install.

Notice: Standalone wrapper installation is unsupported. The operation wrappers depend on their sibling canonical skill; install the complete plugin package.

Since v0.4.0, sepia defines an interface for stacking a voice or style skill on top of it — a minimalism method, a brand voice, a persona guide. It is opt-in: tell sepia the voice skill is in play, and it loads references/voice-skills.md over the normal route. No external voice is loaded unless you say so.

The interface contract operates under fixed precedence rules across operations and routes:

Notice: The voice-skills interface is grounded in one blind review experiment on a strict-minimalism specimen — a worked example, not measured evidence.

Two built-in profiles ship under references/voices/, alongside a persona profile kind with declared override rights:

sepia's rules remove tells. Warmth has another source: the writer knowing who is speaking, and to whom. One case from this project's own release, with nothing measured: the v0.12.0 Threads post was drafted twice by the same model from the same release notes. The first prompt handed it a skeleton (version line, bullets, fixes, engineering note, update line) and got a form filled in. The second prompt handed it an identity, "I maintain this project and just shipped v0.12.0; talk to my followers in my own voice; how you organise it is yours", plus the notes as the only fact source and the platform's limits, and got the post the maintainer kept. A sepia review of that post then found its tells clustered at the opening and the close (an opening about how the change came to be, a maxim ending, a few absolutes), and a refactor took them out without touching the middle.

The order that worked:

1. Identity first.
2. Facts separately, as the only source.
3. Constraints stated as platform limits, with the structure left to the model.
4. sepia as the check.

Do not use tools; everything you need is below.
I am <who>, writing for <whom>, about <what just happened>. Use my own voice; how to organise it is yours.
Platform limits: <length, markup, mentions>.
Facts come only from the notes below; numbers and identifiers must match them.
===== notes =====
<the source document>

A persona profile is that identity written down once so every piece starts from it: what they are to the reader, what their first sentence does, how warmth attaches to a fact, what they never do, and pieces in their own voice as exemplars. The template is references/voices/PERSONA-TEMPLATE.md, the built-in references/voices/personas/nyaneko.md is a worked example, and the interface validator checks the profile structure:

python3 scripts/check_persona.py <file>

Write your own; the exemplars carry more than the description does.

The style pass checks the spread of sentence lengths, the one syntactic measure on which every study that measured it agrees: human text varies more within a passage, in English and in Chinese.

Chinese text loads references/languages/zh.md, calibrating the style pass across four distinct evidence layers:

Notice: The limits of all three empirical Chinese sources are stated in references/languages/zh.md.

Notice: Every command below is written for user scope — install once, use it in every project.

npx skills add Nanako0129/sepia -g     # -g = user scope; the default is project
npx skills update sepia -g             # update
npx skills remove sepia -g             # uninstall

Installs on every agent the Skills CLI supports — Cursor, Cline, Windsurf, Copilot, OpenCode, goose, and more. Pick your agents when prompted. Runtime behavior outside the five platforms below has not been exercised by us; the skill is plain markdown under the Agent Skills standard, so file an issue if your agent trips on it.

Notice: "Verified" means the install completes and the sepia entries appear. The five native plugin installers were each exercised with a live install (QwenPaw's by its contributor). Whether the entries then behave as documented has not been checked platform by platform.

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

# update
claude plugin marketplace update sepia
claude plugin update sepia

Tip: The in-session /plugin install dialog asks you to pick a scope — choose User there.

# install
codex plugin marketplace add Nanako0129/sepia
codex plugin add sepia@sepia

# update — refresh the marketplace snapshot, then re-add to pick up the new version
codex plugin marketplace upgrade sepia
codex plugin add sepia@sepia

# install
grok plugin install Nanako0129/sepia --trust

# update
grok plugin update

Grok also auto-discovers a Claude Code install of sepia if you have one; either route works.

# install directly from GitHub
agy plugin install https://github.com/Nanako0129/sepia

# install: qwenpaw takes a local directory (or a zip URL), so clone first;
# the package's skills symlink resolves inside the clone
git clone https://github.com/Nanako0129/sepia
qwenpaw plugin install ./sepia/.qwenpaw-plugin

# uninstall
qwenpaw plugin uninstall sepia

Notice: Contributor-verified on QwenPaw 2.2.1 (#250, #289; not reproduced by the maintainer): the install completes, the packaged skills symlink is followed into a real tree by shutil.copytree, and /sepia <text> and /sepia-write <text> each start an agent turn through the host's skill dispatch. The other four entries use the same dispatch and were not run one by one. If the package cannot read its skills (a zip install, or a checkout without symlink support), it logs one error and registers nothing; install from a git clone instead.

When one repo should pin its own copy, commit skills/sepia/ into that repo as .agents/skills/sepia (Codex + Antigravity) or .claude/skills/sepia (Claude Code).

Each tool uses its native command:

# Claude Code
claude plugin uninstall sepia@sepia --scope user

# Codex
codex plugin remove sepia@sepia

# Grok Build
grok plugin uninstall sepia

# Antigravity
agy plugin uninstall sepia

# QwenPaw
qwenpaw plugin uninstall sepia

sepia/
├── plugin.json              # Antigravity packaging
├── skills/
│   ├── sepia/                # canonical skill (Agent Skills standard)
│   │   ├── SKILL.md          # routing, operations, calibration rules, guardrails
│   │   └── references/       # passes, rubric, fingerprints, domain rules, languages/zh.md, voice-skills (experimental)
│   ├── sepia-write/SKILL.md  # thin fixed-operation wrappers
│   ├── sepia-review/SKILL.md
│   ├── sepia-refactor/SKILL.md
│   ├── sepia-recreate/SKILL.md
│   └── sepia-hemingway/SKILL.md  # fiction write/refactor with the built-in voice
├── .claude-plugin/          # Claude Code packaging (plugin.json, marketplace.json)
├── .codex-plugin/           # Codex packaging
├── .qwenpaw-plugin/         # QwenPaw packaging (plugin.json, plugin.py, skills symlink)
├── .agents/                 # Codex/Antigravity workspace-mode discovery + Antigravity workflow
└── research/                # digested evidence base with sources

Star history growth over time for Nanako0129/sepia.

Full digests with links are in research/. Primary studies include:

You can use sepia for free without an account. The research behind every rule is open. Ongoing costs are just maintainer time and two kinds of model quota: delegating literature surveys to research agents that read primary papers, and running live models to test rule changes on A/B stories and cross-platform end-to-end reviews before shipping. You can support the project on Patreon.

MIT