Sepia: Making AI Agents Write Like Humans

Where this fits

Your agent drafted an incident postmortem. It is accurate and complete — but it sounds like a language model wrote it: filler density, absent stance, chatbot residue, robotic formatting. You want the facts unchanged, yet the voice human. Sepia is designed for exactly that boundary: it de-AIs writing without losing substance.

Why agents benefit

• Sepia separates diagnosis from editing: the review operation lists AI tells, leaving the text untouched, so the agent can report issues before changing anything.
• Fiction repair works on narrative architecture first, based on StoryScope (arXiv:2604.03136), because style edits barely move a structural classifier.
• Professional prose is routed through domain rules for release notes, announcements, PR and issue replies, code-review comments, incident postmortems, tickets, work orders, technical articles, and blog posts.
• The four operations — write, review, refactor, recreate — give an agent the right amount of intervention, from minimal in-place edits to a full rewrite.
• Every rule is backed by measured human-vs-AI gaps, so the agent is not relying on vibe-based synonym swaps.

Practical scenarios

Incident postmortems

An agent documents a production outage. The draft facts are solid, but the tone slides into corporate filler. Sepia's domain rules restore stance and causality, making the narrative feel written by an engineer who owns the outcome, not by a template.

Code-review comments

An agent generates code-review feedback. Without tuning, comments come out passive and hedgy. Sepia strips the 'it might be worth considering' residue and produces direct, collegial suggestions that a developer would actually leave.

Fiction drafts

An agent writes a short story with mechanical plot symmetry. Sepia's narrative-structure pass repairs pacing and character-driven causality before touching style, avoiding the fake-human slop of simple synonym replacement.

Add it to your agent workflow

Install the skill with:

npx skills add Nanako0129/sepia --skill sepia

Then request one of the four operations explicitly. For example:

Use sepia review on this PR reply. Do not edit; list AI tells only.

Compare before adopting

Before adopting, compare quality signals: does the rewritten text preserve technical details and the author's intent? Does the review catch real tells, not just random word changes? Check maintenance freshness — this is version 0.2.0, so rule coverage will evolve. Compare against generic humanize prompts, which often just shuffle vocabulary, and against style-only editors that miss narrative structure. Finally, assess workflow fit: Sepia pays off most when your agent produces long-form text and needs a final de-AI pass before human review.

Why it is worth tracking

Sepia is worth tracking because it targets the actual root of machine-written prose, not surface symptoms. The structural approach to fiction and the profession-specific routing for office text are both measurable differentiators. With 231 stars, community momentum is real but early. Evaluate it on your own document types now, and keep an eye on updates before you need human-sounding output under deadline. For a live snapshot of the skill and its versions, see https://www.openagentskill.com/skills/nanako0129-sepia.