Skip to content

0xa3b/writing

v1.8.0MIT

Writing skills for prose, technical writing, and agent instructions.

GitHubOpen Website

Writing

This plugin ships writing disciplines for the prose agents produce: a base style for all prose, a layered standard for technical artifacts, and a house style for files that instruct agents.

An agent can invoke all three skills implicitly: prose when a session composes or edits prose, technical-writing when it writes or revises technical artifacts, and agent-instructions when it writes or revises agent-facing instruction files, the scripts a skill bundles, or sub-agent prompts. The skills layer: prose is the base style, and the other two override it for the artifacts they own.

Skills

  • prose: Base style for all prose — plain wording, one main claim per sentence, and a slop-pattern catalog — with a chat-responses reference for conversational replies.
  • technical-writing: Diátaxis document modes, STE-inspired wording, optional controlled-language limits, and Google developer style formatting for technical artifacts — READMEs, guides, runbooks, reference prose, release notes, change descriptions, issue and ticket descriptions, and collaborative comments.
  • agent-instructions: House style and document mechanics for files and prompts that instruct agents — AGENTS.md, CLAUDE.md, SKILL.md, agent definitions, system-prompt fragments, and sub-agent task prompts — and for the scripts a skill bundles.

Prose output style

For Claude Code, select the plugin's prose output style so every chat response applies the prose skill's base style and chat-response rules from the system prompt. Set outputStyle to prose in a settings file, or pick it under Output style in /config. The style keeps Claude Code's software engineering instructions in force.

For chat apps without access to installed skills, such as claude.ai and ChatGPT, paste chatbot-instructions.md into their persistent instructions. On claude.ai, open Settings and paste it into Instructions for Claude. On ChatGPT, paste it into custom instructions. The file stays within the 5,000 characters ChatGPT allows on paid plans; on Free and Go plans, which allow 1,500, paste only the opening paragraph and the sections you want most. From a checkout of this repository, pnpm writing:extract-chat-instructions:copy puts the file on the macOS clipboard.

For agents without output styles, such as Codex, add a pointer line to your user memory (for example AGENTS.md): "Apply the writing:prose skill to chat responses." A similar pointer makes sub-agent dispatch reliable on every agent: "Load the writing:agent-instructions skill before writing a sub-agent prompt."