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).
Table of Contents
Why another humanizer
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:
| Pass | Layer | Examples |
|---|
| 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 |
A 30-feature diagnosis rubric and per-model fingerprints across two layers apply when the writing or executing model is known:
| Model family | Narrative layer (StoryScope) | Sentence-level prose layer (Vendor prompting guides) |
|---|
| Claude | Measured | Claude Fable 5.1 and Mythos 5.1, Fable 5 and Mythos 5, Opus 5, Opus 4.8 |
| GPT | Measured | GPT-5.6, GPT-6 Astra |
| Gemini | Measured | Gemini 3 and 3.1 |
| DeepSeek | Measured | Consulted (no guidance published) |
| Kimi | Measured | Consulted (no guidance published) |
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:
| Domain | The gist |
|---|
| Release notes / announcements | user impact first, artifacts per claim, no marketing inflation |
| PR / issue replies | answer first, cite file:line, no reflex praise, length ∝ 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 |
| Long-form journalism (features, investigations, data stories) | lead and body in two registers, quotations keep their spoken texture, numbers the reader can't size carry a comparison, no summary ending |
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.
Operation entries
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:
This is the beginning of the README. Read the rest at the original. Read the full README ↗
