Sepia De-AI Writing Skill Guide

Key takeaways#

• Sepia is a writing skill for Agent Skills-compatible agents, with integrations named in the source description for Claude Code, Codex, Grok Build, and Antigravity.
• The project focuses on removing obvious AI-writing artifacts and repairing prose structure instead of simply asking a model to rewrite text.
• Builders should treat it as a skill/resource layer: inspect the repository, install it in a compatible agent workflow, and validate output quality against their own style guide.
• The source references StoryScope (arXiv:2604.03136), so teams should separate what the repository implements from what the paper claims.

What Sepia is#

Sepia is a de-AI writing skill for agents. The repository description says it works with any Agent Skills-compatible agent and includes native plugins for Claude Code, Codex, Grok Build, and Antigravity. Its promise is specific: help generated writing sound less like generic model output by applying narrative-architecture repair for fiction and venue-matched rules for professional prose.

That makes Sepia better classified as a resource than a standalone SaaS tool. It is a reusable skill package that can be installed or adapted inside other agent workflows. The practical value is in the rules, examples, and integration pattern.

When to use it#

Use Sepia when a team already has an agent writing workflow and wants a sharper editing pass. It can fit blog drafting, fiction revision, proposal cleanup, internal documentation, and any workflow where the first draft is useful but the final voice needs more human control. It is not a replacement for editorial review. It is a layer that gives the agent a more explicit writing job.

Setup checklist#

1. Open the official repository: https://github.com/Nanako0129/sepia
2. Read the README and installation notes for your target agent client.
3. Confirm that your client supports Agent Skills or the native plugin path you plan to use.
4. Run the skill on a low-risk sample first. Compare the original, model draft, and Sepia-edited output.
5. Add project-specific style rules before using it for public copy.

Evaluation criteria#

The right test is not whether Sepia makes text longer or more polished. The right test is whether it removes repetitive model patterns, preserves facts, improves structure, and keeps the desired voice. For professional prose, check whether venue-specific rules actually match the target publication. For fiction, check whether scene logic, rhythm, and character intent improve without flattening the style.

Builder cautions#

Any rewriting skill can change meaning. Teams should keep source drafts, require review for factual claims, and avoid running automatic edits on legal, medical, or high-stakes content without human approval. If Sepia is used inside Claude Code, Codex, or another agent loop, scope file permissions carefully so the skill cannot rewrite unrelated project files.

Source#

Primary source: Nanako0129/sepia. Repository description captured by the OpenTools triage pipeline: De-AI writing skill for any Agent Skills-compatible agent (77+ via the Skills CLI), with native plugins for Claude Code, Codex, Grok Build, and Antigravity. Narrative-architecture repair for fiction, venue-matched rules for professional prose. Based on StoryScope (arXiv:2604.03136).