Agent Skill 'sepia' Released to Correct Narrative Structure of AI-Generated Text

The agent skill sepia has been released to improve Narrative Architecture beyond superficial vocabulary correction to erase traces of AI-generated text. Unlike existing tools that focused on correcting translation-style issues at the sentence or vocabulary level, sepia prioritizes resolving discourse-level problems such as causal relationships and conclusion structures.

How It Works and Features

• Narrative-First Principle: Based on research from the paper StoryScope, released in April 2024, it addresses the finding that AI detection performance does not significantly degrade when only surface style is modified. It adopts an approach where narrative structure is corrected first, and surface style is modified later.
• 4 Fixed Operations: Provides four modes: write (new writing), review (diagnosis), refactor (minimal modification), and recreate (rewriting keeping only facts).
• Diverse Document Support: Includes rules for five types of practical documents, including novels, essays, release notes, PR responses, and incident postmortems.
• Safety and Transparency: Does not fabricate specific content and limits AI model estimation features. It clearly documents the distinction between design reasoning and actually measured evidence.

Installation and Compatibility

It follows the Vercel Skills standard and can be installed as a plugin in major AI coding agent environments such as Claude Code, Codex, Grok Build, and Antigravity. Distributed under the MIT license, it allows for both personal and commercial use.