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, and Antigravity 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

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

Plus a 30-feature diagnosis rubric and per-model fingerprints in two layers: narrative tells measured by StoryScope (Claude, GPT, Gemini, DeepSeek, Kimi) and sentence-level tells taken from the vendors' own prompting guides (Claude Fable 5.1 and Mythos 5.1, Fable 5 and Mythos 5, Opus 5, Opus 4.8; GPT-5.6, GPT-6 Astra; Gemini 3 series), applied when the writing or executing model is known. Vendors that publish no such guidance are recorded as consulted, not guessed.

Professional prose fails differently. 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

The governing principle throughout: 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

---

GitHub에서 전체 내용 보기