Skip to content

aieda1l/proseshape

v1.4.2MIT

Rewrite or write prose so it reads as chosen by a person, not defaulted by a model, while keeping every fact, quotation, plot event, and the writer's voice.

GitHubOpen Website

ProseShape

Make prose feel chosen, not defaulted.

ProseShape is a writing skill for Claude and Codex. It rewrites or writes prose so that each choice reads as made by a person for this reader, not inherited from a model's defaults, while keeping every fact, number, quotation, citation, plot event, and the writer's voice.

It works at two levels:

  • Sentences. It removes the habits that make text read as generated: staged contrasts, one-line closers, inflated significance and chatbot residue. Then it puts the life back in through the writer's own material.
  • Story structure. It looks for the narrative defaults that research on AI fiction identifies: stated morals, tidy single-track plots, emotions told through the body, and epiphany or epilogue endings.

What it does

  • Rewrites supplied text and keeps every claim, event and ending. It flattens nothing and invents nothing.
  • Light edits. Already-good writing gets a careful copyedit or is left nearly unchanged.
  • Restructures fiction when you give it free rein, and reports each structural change so you can reject it.
  • Matches your voice from a sample of your own writing.
  • Writes from scratch, choosing against the default version of the piece instead of cleaning it up afterward.
  • Keeps technical text exact. Code, commands, terms, numbers and citations stay unchanged.

It is built for readers, not AI detectors. It does not try to evade detection, and it does not add typos or fake quirks.

How to use it

Ask for what you want and paste the text, for example:

  • "Humanize this story. Same events, same ending."
  • "Light polish only. Keep every fact and citation exactly as is."
  • "Here are two things I wrote. Rewrite this draft so it sounds like me."
  • "The prose is fine but this story feels AI-shaped. Free rein to restructure; keep the characters and premise."
  • "Write a 1,200-word story about a hospice night nurse. No stated moral, no years-later ending."

Claude and Codex use the skill when a request matches. To be sure they do, name it ("use ProseShape"). In Claude Code you can also run /proseshape:proseshape; in Codex, pick it from /skills.

What is inside

ProseShape is instructions only: a SKILL.md, six reference files, and the manifests and icon that Claude and Codex read. It runs no code, starts no servers, fetches nothing, and sends nothing anywhere.

Evidence

The skill was developed and tested against a strong ordinary-editor prompt, using held-out texts and a blind judge. On 16 texts it had never been tuned on, version 1.4.2 scored 8.02 out of 10:

  • The editor prompt scored 7.44, a lead of +0.58 (95% interval +0.19 to +1.02).
  • Version 1.4.0 scored 7.86. The gain over it is small (+0.16).
  • Fact problems. The judge found 2 in its output, against 19 for 1.4.0 and 37 for the editor prompt.

The method, raw outputs and limitations are in the source repository.

License and credits

MIT. ProseShape adapts material from Humanizer by Siqi Chen (MIT), and an idea from HumanScope (MIT). It draws on Wikipedia's Signs of AI writing (CC BY-SA 4.0) and on StoryScope: Investigating idiosyncrasies in AI fiction by Russell et al. (arXiv:2604.03136). See THIRD_PARTY_NOTICES.md.