Agents

Working on an Ascribe project with an AI coding agent, such as Claude Code, GitHub Copilot, or Cursor, and handing it the problems Ascribe finds.

An agent is the AI coding tool you work with: Claude Code, GitHub Copilot, Cursor, Codex, and others. Ascribe gives your agent what it knows about a project, in a form an agent reads well, and checks what the agent writes. Its tools answer questions and return edits; your agent makes them, with your approvals.

Nothing here is set up until you run a command for it. The quickest start, in a project’s folder:

ascribe agents sync

That writes AGENTS.md and the skill, which most agents read on their own. Then add what your agent can use:

• Work with an agent: the loop of editing a page and checking it, and the commands that answer what an agent would otherwise guess.
• Write the agent files: your project’s rules in AGENTS.md, and the skill that teaches the loop.
• Prompt your agent: Prompt agent on a problem, a review comment, or a changed page.
• Set up the MCP server: the same answers as tools, for an agent without a shell.
• Check an agent’s edits with hooks: a check it can’t forget, after each edit and before it finishes.
• Claude Code and GitHub Copilot: what each needs beyond the rest. In VS Code, the extension sets up most of it: see Agents in the editor.
• Index your docs for your product’s agents: for the agents of the people who use your product, not the ones that write your docs.

What’s sent where says what leaves your machine, and what to watch for when your agent reads someone else’s text.

This guide writes commands as ascribe. With @ascribed/cli installed in your project, as Install says, run them as npx ascribe (pnpm exec ascribe with pnpm). The files sync writes tell your agent to do the same: see The command they name.