sepia

sepia is a writing skill that attacks AI tells at the structural layer instead of the word layer. Most humanizers swap vocabulary and shuffle syntax. sepia cites StoryScope, a 2026 study of 61,608 stories, which found a classifier using narrative-structure features alone identifies AI fiction at 93.2% macro-F1 and barely drops when surface style is edited away. So it works on the architecture: narrators who explain the theme, causally tidy single-track plots, emotion rendered only as bodily sensation, no real-world references, endings resolved by growth and acceptance. MIT licensed.

It installs as a portable Agent Skill, so any agent speaking that spec can load it, with native plugin packaging verified on Claude Code, Codex, Grok Build and Antigravity. One canonical SKILL.md, no per-platform forks. Four operations: write, review (diagnose only), refactor (minimal edits), and recreate (full rewrite). Professional prose gets its own rule files per venue, covering release notes, PR replies, postmortems, tickets and technical articles.

Free with no tiers. Anyone whose drafts come out sounding like a model wrote them gets real value, especially on the professional side where the rules are concrete: answer first, cite file and line, no reflex praise, length proportional to stakes.

The catch is that this is a prompt, not a guarantee. It calibrates toward the human distribution rather than inverting the AI one, which is the right instinct, but nothing here beats an actual detector and treating it as a compliance tool would be a mistake.