Skills

sepia: A De-AI Writing Skill Across Claude Code, Codex, Grok Build, and Antigravity

sepia is best for removing the "AI flavor" from AI-generated text in Claude Code, Codex, Grok Build, and Antigravity — fiction first fixes narrative architecture, professional documents follow register-matched rules; the most important thing to know before use is that it installs via each platform's native installer or Antigravity's manual path, and different text types follow different flows.

Official source · Nanako0129/sepia

Install

claude plugin marketplace add Nanako0129/sepia && claude plugin install sepia@sepia --scope user

Who It's For

  • Writers who need to remove the "AI flavor" from AI-generated text: this includes fiction authors writing novels, who need to fix narrative architecture before polishing wording, as well as people writing professional documents such as release notes, PR replies, postmortems, tickets, and technical articles.
  • Users of Claude Code, Codex, Grok Build, or Antigravity; the skill is a single SKILL.md that works across all four clients, with no need to maintain separate versions per platform.
  • Users who want to install once and have it available in all projects: the official default installation scope is user scope.
  • When it's not a fit: the official evidence does not state which text languages it supports, nor whether it works with clients beyond the four listed above; this information is unknown, so don't assume it applies to other agents or arbitrary languages. It addresses the quality of text output; the evidence does not say it replaces fact-checking or content-accuracy review.

First Try in 3 Minutes

The recommended main path is installing via Claude Code's native plugin marketplace (official commands, user scope):

claude plugin marketplace add Nanako0129/sepia
claude plugin install sepia@sepia --scope user

Success signal: the plugin installs without errors; afterwards you can invoke this skill in any project within that client.

For your first attempt, pick a low-risk task: run a review diagnosis with the agent — for example, paste in a piece of release notes or a PR reply you've already written and suspect has AI traces, and ask the agent to diagnose it using sepia's review operation. Note: the official documentation only says review is a "diagnose only" operation, which is enough to show its output is a diagnosis, but the evidence does not state whether review itself reads or writes files, executes commands, or accesses the network — these behaviors are unknown and cannot be treated as zero-risk. Review is a diagnosis and does not directly rewrite your text.

On success you'll observe: the agent gives diagnostic feedback on this professional document according to sepia's rules; sepia applies register-matched rules to professional documents (release notes, PR replies, postmortems, tickets, technical articles). If you provide fiction text, the flow is different: fiction first repairs the narrative architecture, then moves to the wording level — choose the task according to your text type.

Other official installation methods (secondary notes, not parallel main entry points):

  • Codex: codex plugin marketplace add Nanako0129/sepia then codex plugin add sepia@sepia
  • Grok Build: grok plugin install Nanako0129/sepia --trust
  • Antigravity: no marketplace; the official path is manual — clone release v0.2.0 to $HOME/.sepia, then copy the skill to $HOME/.gemini/config/skills/sepia and the workflow to $HOME/.gemini/antigravity/global_workflows/sepia.md; it aborts if the target already exists.
  • Skills CLI: npx skills add Nanako0129/sepia -g (-g means user scope; the default is project).

Use Cases

  1. Diagnosing AI traces in a PR reply: scenario — you had AI help write a PR reply and worry the tone is mechanical. Send the text to the agent and ask for sepia's review operation. Expected result: diagnostic feedback tailored to PR replies as a professional document, following its register rules, with no rewriting.
  2. Rewriting a postmortem or technical article: scenario — you have a draft that needs the AI flavor removed with minimal changes. Run a refactor task with the agent (officially defined as minimal edits). Expected result: a revised version within the scope of minimal changes; it rewrites the text content you provide — the evidence does not say the skill directly modifies files on your disk, so it's best to paste the text in and save it yourself.
  3. Fixing a novel chapter: scenario — an AI-generated chapter doesn't read like human writing. Run a recreate task (officially defined as full rewrite) and hand the chapter text to the agent. Expected result: a rewritten version that first repairs the narrative architecture and then handles wording, per sepia's flow. This is an output action with side effects (it generates a large amount of rewritten text), so confirm you've provided the correct chapter before starting.

Usage Boundaries

  • Installation differences must be considered per platform: Claude Code, Codex, and Grok Build each use their own native plugin installers; Antigravity has no marketplace and only supports the official manual path (pinned to v0.2.0); Skills CLI is another official alternative, defaulting to project scope, with -g for user scope. A project-level pinned copy can be committed as skills/sepia/ at .agents/skills/sepia (Codex + Antigravity) or .claude/skills/sepia (Claude Code).
  • Uninstalling: Claude Code uses claude plugin uninstall sepia@sepia --scope user; Codex uses codex plugin remove sepia@sepia; Grok Build uses grok plugin uninstall sepia; Antigravity requires renaming both the skill and workflow entries with a .disabled suffix (the official precheck aborts if the source is missing or a .disabled target already exists), and $HOME/.sepia is kept for inspection — deleting it is a separate manual decision.
  • Updating: Claude Code uses claude plugin marketplace update sepia + claude plugin update sepia; Codex requires codex plugin marketplace upgrade sepia followed by re-running codex plugin add sepia@sepia; Grok Build uses grok plugin update; Skills CLI uses npx skills update -g.
  • Unconfirmed information: whether it reads or writes files, executes shell commands, or accesses the network; which text languages it supports; whether it works with clients beyond the four listed. None of this is stated in the evidence — if sensitive files or network environments are involved, verify the behavior yourself before use.