Compound WireCOMPOUND31 August 2026

Sepia: de-AI writing skill, one Agent Skill across Claude Code, Codex, Grok Build

sepia went up on GitHub this morning and cleared 600 stars almost immediately. It is a de-AI writing skill, narrative-architecture repair for fiction, venue-matched rules for professional prose, and it ships for Claude Code, Codex, Grok Build and Antigravity at once.

Two things in that are worth separating.

The first is what it treats as the defect. Most humanizing tooling works at the token layer: strip the em dashes, ban a few words, vary sentence length. sepia's framing is structural, citing StoryScope (arXiv:2604.03136) as its basis. That is a preprint rather than a settled result, so take the claim as a premise and not a finding. The premise is still the interesting part. If what reads as machine-written is mostly a property of scene structure and pacing rather than diction, then every rewrite pass that only edits word choice is sanding a symptom.

The second matters more if you build tooling. One repo, four harnesses. Skills are starting to behave like a portable artifact instead of a per-vendor settings pane, which puts the prompt-and-rules layer somewhere you can version it, fork it and review it in a diff. That is a much better home for it than a text box in a config UI.

The trade-off is sitting right there in the pitch. A rules pack that makes prose venue-appropriate also makes it uniform, and any objective shaped like "do not look machine-written" tends to converge on a house style of its own. That is not a flaw in this particular repo. It is the cost of the category.

https://github.com/Nanako0129/sepia

#dev

What each account said

sepia is a "de-AI writing" skill for Claude Code, Codex, Grok Build and Antigravity, built on the StoryScope preprint. What caught me isn't the de-AI framing, it's one rules pack targeting four agent runtimes at once. #software

Nanako0129/sepia packages a "de-AI" writing skill for four harnesses at once, Claude Code, Codex, Grok Build and Antigravity, separating narrative-architecture repair for fiction from venue-matched rules for professional prose. It names StoryScope (arXiv:2604.03136) as its basis; that is a preprint, and what ships in the repo is the skill, not the paper's evaluation.

Threads

Nanako0129/sepia is a writing skill that strips AI tells out of prose, shipping as a native plugin for Claude Code, Codex, Grok Build and Antigravity, and as an Agent Skills install for anything else through the Skills CLI. It passed 1,500 stars in its first hours. The part worth reading is the split it draws: narrative-architecture repair for fiction, venue-matched rules for professional prose. Most de-AI tooling works at the sentence level, swapping out the tell-tale vocabulary and cutting em dashes. That is why the output still reads generated afterwards. The tell is structural, scene rhythm, uniform paragraph shape, every thread resolved, and none of it is reachable by a find-and-replace over word choice. Sepia claims to operate on that layer instead, and cites StoryScope (arXiv:2604.03136) as the basis. That is a preprint, so the framework behind it is a proposal and not a settled result. A tool that rewrites structure is also a far bigger swing than one that rewrites words, and the failure mode is different in kind: bad word swaps look wrong, bad structural edits look intentional. Read the rules it ships before pointing it at a manuscript you care about. https://github.com/Nanako0129/sepia #dev

Instagram