sepia skill
A de-AI writing skill: it repairs the narrative architecture of fiction and edits professional documents by their venue rules
- 3,798 ★ on the repository
- Last commit
- MIT
Minimal access
Why this level
- The skill consists only of instructions and reference material and runs no code
Install
- Works in
Manual install
npx skills add Nanako0129/sepia -gInstall via the Skills CLI onto the agents you pick; the -g flag installs at user scope. Update with npx skills update sepia -g.
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- The skill consists only of instructions and reference material and runs no code
This is third-party code. Review the repository files before installing.
What it does
The skill turns model-generated text into prose that reads as written by a person. For fiction it works at the level of narrative structure, not just words: it removes the narrator explaining the theme, single-track plots, emotion rendered only as bodily sensation and other tells that give a machine away. Professional text is routed by its venue: release notes, pull request replies, incident postmortems, tickets, technical articles, long-form journalism. Four modes are available: write, review without edits, minimal in-place fix, and full rewrite. The rules draw on research into text generation, with debatable calls marked as editorial assumptions.
Who it is for. For writers, editors, developers and communications people who need text without traces of machine generation.
Good fit when
- You need to strip AI tells from an article or story while keeping the author's voice
- You need a diagnosis of a text with a list of weak spots and no edits
- You need to write or rewrite a professional document by its venue rules
Not a fit when
- You need fact-checking rather than the form and style of the text
- You want to game an AI-text detector rather than improve readability: the skill is tuned for a human reader, not for fooling detectors
Example request
Strip the AI tells from this text and keep the author's voiceLimitations
The skill checks the form and structure of a text, not the truth of its facts, and it is explicitly forbidden to invent data. Some professional-text rules are editorial assumptions rather than measured findings, and this is marked in the skill itself. Chinese text loads a separate calibration. The skill works only where the agent supports the Agent Skills standard.
How to disable. Remove the sepia plugin with your agent's own command, for example claude plugin uninstall sepia@sepia, or delete the skill folder from the skills directory.
FAQ
Is this another AI-detector bypass?
No. The skill is tuned for a human reader and does not try to fit automated detectors. The goal is text that reads naturally, not text that passes a specific check.
How do the modes differ?
Review only diagnoses and changes nothing. Refactor makes minimal in-place edits. Recreate rebuilds the text from the facts and intent of the original.
Discussion
No questions or reviews yet.
Related
SkillsA skill that strips AI writing tells from text using Wikipedia's list, without adding invented facts
A skill that removes AI tone from work writing while keeping every fact, commitment and caveat at the same strength.
Avoid AI Writing
avoid-ai-writing
A skill that finds typical machine-writing patterns in text and rewrites them, with detect-only and edit-in-place modes
A skill that finds AI writing tells, rewrites the text in a chosen voice, and optionally scores it from 0 to 100
A skill that edits English prose: it strips filler and cliches, keeps facts and voice and marks gaps instead of inventing.
A Russian text editing skill: removes bureaucratese, calques and AI mannerisms without touching facts, with a scanner to measure the edit
README in short
The README explains that ordinary text cleaners edit only words, while sepia works at the level of narrative structure, where AI tells are most visible. For fiction it runs three passes: narrative architecture, discourse flow and surface style, plus a diagnostic rubric and per-model fingerprints. Professional text follows thin per-venue rule files on top of one shared checklist. The skill installs through the Skills CLI, which supports more than 77 agents, and as separate native plugins for Claude Code, Codex, Grok Build, Antigravity and QwenPaw. The full evidence base with links to studies lives in the research folder. MIT licensed.
README badge
Original SKILL.md
--- name: sepia description: Make AI-generated writing read as human-written, in fiction and in professional prose. Repairs the narrative architecture of fiction and stories (based on StoryScope, arXiv:2604.03136); routes professional text through domain rules for release notes, announcements, PR and issue replies, code-review comments, incident postmortems, tickets, work orders, technical articles, blog posts, and long-form journalism. Four operations - write, review (diagnose AI tells without editing), refactor (minimal in-place edits), recreate (full rewrite). Use when asked to humanize, de-AI, unslop, or strip AI flavor from any text; when writing or revising any of these document types; or whenever output must not read as machine-written. license: MIT metadata: version: "0.10.0" --- # Sepia — de-AI writing This skill combines measured findings with marked editorial heuristics. In fiction, StoryScope's narrative-only classifier reached 93.2% macro-F1, while its Core Only 30-feature XGBoost held-out classifier reached 84.8% macro-F1 (AUPRC .828); the manual rubric is neither classifier. The professional path combines measured studies with editorial heuristics, and its prescriptions are Sepia inferences unless a source explicitly tested the intervention. Route first, then operate. Sepia writes for expert human readers and is tuned to pass no automated AI-text detector.