Your AI Writing Still Sounds Like AI. Here's Why.

Most tools that claim to "humanize" AI writing do the same thing: swap a few words, break up a sentence, trade a semicolon for a period. A study published this year says that layer barely matters. The flatness a reader actually feels lives one level deeper, in how the story is built.

The research behind it

StoryScope (Russell et al., 2026) studied 61,608 stories written by humans and five frontier models: Claude, GPT, Gemini, DeepSeek, and Kimi. Looking only at narrative-structure features, with no wording signal at all, AI-written fiction separated cleanly from human writing at 93.2% macro-F1.

Then the researchers rewrote the AI-sounding wording away by hand and re-measured.

The separation barely moved: 95.5% down to 93.9%.

The structure was doing almost all the work, not the vocabulary.

What actually reads as synthetic

The features that survive a wording rewrite are structural, not lexical. This is the kind of thing an editor flags, not a thesaurus:

• The narrator explains the theme directly, instead of letting it emerge
• The plot resolves on one tidy, causally-clean track
• Emotion shows up only as physical sensation, a racing heart, a tight chest
• Nothing references the real world outside the story
• The ending resolves too neatly, usually through growth and acceptance

None of that is a sentence-level habit. It is architecture: the shape of the story before a single word gets chosen. Fix the shape, and the writing gets better whether or not anyone is running a classifier over it.

sepia works on that layer first

A new open-source Agent Skill, sepia, is built directly on the StoryScope findings, with citations to eleven related studies in its own repo. It hit #2 trending on GitHub within days. It runs in three passes, in this order:

1. Narrative architecture (fiction)
   loosen the causal chain, delay reveals, mix how emotion is shown,
   thin out the character network, name real things

2. Discourse flow
   remove the repeated paragraph-question pattern, fix the sagging
   middle, vary rhythm and sentence position

3. Surface style
   the layer every other humanizer stops at: clichés, syntax
   templates, word choice, register

It also runs as a pure diagnostic. Paste in writing and it flags what reads as flat or synthetic without touching a word of it. For professional documents (release notes, PR replies, postmortems, tickets, technical articles) it swaps in a separate rule set per document type, matched to what that document actually needs to do its job.

The part I trust most: it doesn't max out every rule

The interesting design choice is restraint. A story that applies every available fix becomes its own fingerprint. Real writing sits at moderate values, not at the extreme in either direction. sepia picks three to five techniques per piece and leaves the rest alone, on purpose. That is a craft decision, not a checklist rule.

Install through the Skills CLI, works with Claude Code, Codex, Grok Build, Antigravity, and 77+ agents:

npx skills add Nanako0129/sepia -g

MIT licensed. 1,860 stars as of this week. Real research citations sit in the repo's own research/ folder.