De-AIing text by fixing the blueprint, not just the wallpaper

What it does

sepia is a portable Agent Skill for Claude Code, Codex, Grok Build, and Antigravity that makes AI-generated writing harder to spot. For fiction, it runs a three-pass protocol—narrative architecture, discourse flow, then surface style—backed by a 30-feature diagnosis rubric and per-model fingerprint corrections. For professional prose, it applies venue-matched rules to release notes, PR replies, postmortems, tickets, and technical articles.

The interesting bit

Most “humanizers” swap synonyms and shuffle syntax. The StoryScope study (61,608 stories) showed that narrative-structure features alone detect AI fiction at 93.2% macro-F1, while surface edits barely dent detection. sepia targets those architectural tells—causally-tidy single-track plots, narrator-explained themes, emotions rendered only as bodily sensation—rather than polishing prose.

Key highlights

• One canonical SKILL.md works across four agent platforms with no per-platform forks
• Four operations: write, review (diagnose only), refactor (minimal edits), and recreate (full rewrite)
• Fiction protocol selects only 3–5 structural moves per story to avoid creating a new “perfect human” fingerprint
• Professional rules are domain-specific: release notes lead with user impact, postmortems are blameless toward people but merciless toward mechanisms
• Governing principle: calibrate to the human distribution, don’t invert the AI one

Caveats

• The README is thorough on mechanics but offers no before/after examples of the fiction or prose output
• Antigravity support requires a manual install path with no automated updater

Verdict

Worth a look if you publish fiction or professional docs through an AI assistant and want the output to escape the usual structural tells. Skip it if you just need a thesaurus with a CLI.