Proof

This lets Claude read and write collaborative markdown documents through Proof, either via web API or by driving the local macOS app. Every edit gets attributed to a stable agent ID (ai:compound-engineering by default), so humans can see what the AI changed. The real workflow here is human-in-the-loop review: your agent pushes a draft to Proof, collects structured feedback as inline marks (changes, questions, objections), then syncs the final doc back to disk. It's built for the moment when you need more than a diff, when you want threaded comments and presence indicators before committing generated content. If you're shipping agent-written docs that need approval, this bridges the gap between generation and merge.

Install to Claude Code

npx -y skills add everyinc/compound-engineering-plugin --skill proof --agent claude-code