Content Creation

How jzOcb Built a Writing Style Skill That Learns from Your Edits

Built by jzOcb · · Source: GitHub

Quick answer

jzOcb published an open source writing style skill template for Claude Code and OpenClaw. The AI drafts from SKILL.md, you edit the draft, and a script compares the two versions, extracts style rules and writes them back into SKILL.md.

Result: 274 GitHub stars and 27 forks on the repo when we read it. The README gives no measured results beyond saying each round makes the next draft more accurate.

What they built

jzOcb published an open source template for a writing style skill with auto-learning built in. The skill is a SKILL.md file that tells the AI how you write. What makes it different is that the file updates itself based on the edits you make to the AI's drafts. The repo is MIT licensed, and the README (written in Chinese) is the only source here. It describes the tool and does not include a personal story from the author.

How it works

According to the README, the loop is: the AI writes a first draft using SKILL.md, you edit it until you are happy, the two versions are compared, rules are extracted, and SKILL.md is updated so the next draft is closer. It only needs two data points, the AI's original draft and your final version. The README says the number of editing rounds in between doesn't matter.

  • Start with SKILL.md. You can write your own style rules into it, or leave it blank and let the auto-learning fill it in.
  • Record both versions. A script called observe.py saves the original draft and then your final version. The README describes it as pure Python with no dependencies.
  • Extract the rules. A second script, improve.py, sends the difference to an LLM, which turns it into writing rules. Rules are ranked P0, P1 or P2 by confidence, and P0 rules are applied automatically.
  • Backups. The README says a backup is made before every update and that rolling back takes one step.

The improve script needs an LLM command line tool. It looks for the claude command from Claude Code first, then the llm tool, or you can point it at a custom command with the IMPROVE_LLM_CMD environment variable.

For Claude Code, the install is to clone the repo and copy it into your skills folder. For OpenClaw, the README gives a ClawHub install command.

The result

The README does not report any measured results. Its claim is that each round of edits makes the next draft more accurate. The repo page showed 274 stars and 27 forks when we read it.

How to build something similar

  • Keep your style guide in one plain text file that the agent reads every time it writes for you.
  • Save the AI's untouched draft before you start editing, and save your final version when you finish. Those two files are the training data.
  • Ask a model to compare the two and propose rules, then only auto-apply the ones it is most sure about. Review the rest by hand.
  • Back up the style file before each change so one bad rule is easy to undo.

The original post

jzOcb · View on GitHub

Want to build something like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week.

Join Shipping Skool — Free

More use cases

← All AI agent use cases