sepia

sepia is a writing skill for coding agents, and its argument is about where the tell actually lives: in structure rather than in word choice. Every popular humanizer edits vocabulary and syntax; this one starts from a study in which human editors rewrote the surface style of AI fiction and a classifier still caught it — detection fell only from 95.5% to 93.9% — so fiction goes through a narrative-architecture pass before anything touches the prose, and each professional document type gets a thin rule file matched to its venue on top of a shared checklist. Four operations — write, review, refactor, recreate — run on one canonical SKILL.md, installed by any agent that speaks the Agent Skills standard (the Skills CLI claims 77+) and packaged as a native plugin for Claude Code, Codex, Grok Build, Antigravity and QwenPaw. What it does not know is labelled rather than filled in: the evidence behind each rule sits in a research directory with fifteen primary sources, and vendors that publish no prompt guidance are recorded as consulted instead of guessed.

Who built itThe repository’s 303 commits come from six accounts: 283 of them from the maintainer’s own, under two commit names — Nanako Tsai on 214 and Nyanako on 69, both mapped to the same GitHub account — then AugustusW 13, CallMeHFK 2, shihyuho 2, Franky100-pig 1 and javaht 1, with one commit left unlinked. Two thirds of the history, 202 commits, carries a co-author trailer, and 201 of those name a Claude model: Fable 5.1 on 152, Fable 5 on 25, Opus 5 (1M context) on 19 and Opus 5.5 (1M context) on 5; the one remaining trailer names a person.

1. 01

   Twenty-seven days, 303 commits, and what the trailers say

   The repository was created on 2026-08-28 and its last push is 2026-09-23: twenty-seven days, 303 commits and fourteen releases, from v0.2.0 on the day it appeared to v0.12.2 on 2026-09-22. The pace is uneven — 45 commits in August, 258 in September — but the shape of the authorship is what makes the history worth reading. 283 of the 303 commits come from one account under two commit names, Nanako Tsai on 214 and Nyanako on 69; 202 carry a co-author trailer, and 201 of those name a Claude model — Fable 5.1 on 152, Fable 5 on 25, Opus 5 (1M context) on 19, Opus 5.5 (1M context) on 5 — against a single trailer naming a person. Five other accounts contribute the rest: AugustusW 13, CallMeHFK 2, shihyuho 2, Franky100-pig 1, javaht 1, and one commit carries no linked account at all. The release pull requests are written in the same dialect as the code: each is a version bump of four declarations checked by python3 scripts/check_versions.py, each closes with a Claude Code session link, and each says outright that release notes live on the tag rather than in the body. The documentation was produced the same way, and by the project itself: the three READMEs were restructured by the Antigravity CLI running sepia’s own recreate operation on the documentation route, with the current file as the only fact source, the style guide inlined, and languages/zh.md loaded for the two translations.

2. 02

   The research directory is the product, and it carries its own limits

   What separates this from a prompt pack is one habit: every rule names its source, and every source names what it does not cover. research/sources.md is 61,944 bytes — the largest file in a repository that is mostly Markdown — and the README’s source table lists fifteen primary studies. The founding one is StoryScope: 61,608 stories by humans and five frontier models, where a classifier using narrative-structure features alone reached 93.2% macro-F1 on AI fiction. The number that decides the design is the next one: in the same study’s edited condition, where human editors rewrote the surface style, detection moved only from 95.5% to 93.9%. That gap is the argument for repairing architecture instead of vocabulary. The professional route then ran on that argument with no measurement of its own, and issue #263 says so plainly — on fiction the claim carries a figure, on professional prose it did not. A preprint published on 2026-09-14 closed part of the gap: 2,250 pre-ChatGPT human blog posts from 268 company domains against 11,250 AI mirrors from five frontier models, 214 features across 11 dimensions of which 187 are structural, reaching 98.0 macro-F1 from structural features alone on held-out companies and 98.1 after each model reworded its own post. It entered the ledger as SLOPSHAPE-2026 with its limits attached: the features are LLM-scored, and the experiment never tested human editing.

3. 03

   How the claim is checked, and the four days nobody noticed

   The distance between what the skill claims and what the repository measures is visible, and reported. There is one behavioural eval — evals/deaify-release-note/, a 1,160-byte prompt and three graders of 104, 234 and 889 bytes, asking whether the skill fired, whether the output reads human and whether it carries slop markers — and it runs on every push through claude plugin eval. That workflow was red for four days and the failure looked like nothing: from 2026-09-15, seven consecutive runs (34980491671, 35005656553, 35107215014, 35119480316, 35368600325, 35368661029, 35368703581) exited 1 before the first case ran, because the CLI had gained a first-run trust prompt that a runner cannot answer. The last green run was 34592097838 on 2026-09-11. Issue #259 lists the seven ids and notes that the exit is immediate, so the summarise step printed nothing; #260 adds --trust-plugin with a comment stating what the flag asserts and what it does not. Almost everything else that checks this repository is text and Python: two validators, 88 tests when v0.11.0 shipped on 2026-09-18 and 94 by 2026-09-20; a CodeRabbit configuration whose auto_pause_after_reviewed_commits was raised from 5 to 20, because at the default a paused review is indistinguishable from a slow one; and live-model A/B runs, which the README lists as an ongoing cost of the project rather than a feature of it.

4. 04

   The persona turn: measurements that made the writing read as a format

   The sharpest reversal in the repository is about voice. Issue #257 opened as an RFC with evidence behind it: a private pilot from 2026-09-16 to 2026-09-18 had produced nineteen persona profiles, each written from close readings of ten to fifteen articles by one writer rather than from metrics, and every one of the nineteen needed at least one existing rule to yield. Step two shipped the interface, the template and scripts/check_persona.py; step three ran it end to end, and the maintainer’s report lists what held — the route gate, the persona-cost block, the categories that never yield — beside five gaps, among them a Consent: field with no honest value for a privately held profile. Step four then rewrote the first built-in persona twice. Five runs on one fact list settled it: the bodies built on sentence-length shares, emoji density and k over n counts read as a format rather than a person, and the output a reader recognised came from the writer’s own prose specification, which carries no distribution target. v0.12.0 carried it into the interface — the template became prose-first with fifteen fixed sections and one optional exemplars section, the validator now requires a blind-test record of none yet under Tested: untested, and a persona written to the previous template no longer validates, which is why a 0.x release took a minor bump rather than a patch.

5. 05

   What other people sent, and what the project would not do

   Thirty issues and pull requests sit in the report; the most instructive are the ones the project did not take. Issue #268, filed by a contributor, is a limitation stated against the project’s own interest: sepia’s output can still be flagged as mixed by detectors, because it deliberately does not optimise for burstiness or perplexity — the two axes those detectors lean on — and because it states that it is not trying to evade them. The same contributor then built the optional skills that would have covered those axes; the first pull request was closed in favour of a second whose title, body and branch name never mention a detector, on the argument that the wording inverts the project’s stated position. Issue #267 is a thank-you from someone maintaining manuscript-polishing skills for journal submission, and it is careful about the boundary: the framing was taken, no text, no rule files, no corpus. Issue #274 measures a shipped package — an install whose skills symlink has been flattened reports success and lands nothing — and its author posted a correction retracting the fix he had proposed first. Issue #283, the newest item in the report, is the least flattering: a 2,900-character Chinese popular-science essay falls through the routing table into the generic prose route, so the narrative and discourse passes, where its defects actually are, never run.

6. 06

   A fingerprint table kept honest one vendor page at a time

   A 19,021-byte reference file claims to know how each model writes by default, and the pull requests are the record of what it costs to keep that claim honest. When Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna shipped, #275 recorded them as consulted with no prose statement and gave them no operative row, leaving the existing tables in place as priors, because their pages say nothing about default length, register, formatting or tone. #277 moved the Gemini row down from an upper bound the vendor never wrote to the scope the pages actually support — Gemini 3 and 3.1 operative, 3.5 and 3.8 Flash priors only — on pages re-read on 2026-09-23 after the vendor had updated them on 2026-09-17. #280 re-read the Opus 5.5 page, found two more statements about writing, and replaced a claim that there were only three with all five quoted, next to the reason the conclusion still holds: none of the five names a default length, register, formatting or tone. The newest is an audit of the skill’s own text by a vendor tool — #281 ran Anthropic’s /claude-api prompt-audit, the spec bundled with Claude Code 2.1.280 and byte-identical to an open pull request in anthropics/skills at a named head commit, applied six medium-confidence rewrites and left routing, operations and rules unchanged. One of the six makes a citation name GPT-5, DeepSeek-V3 and o3-mini and admit that no Claude model was measured.