I use multiple AI coding tools. Claude Code, Cursor, OpenCode. They each give access to different models, and I switch between them depending on what I need. I want them all to follow the same project instructions. That shouldn’t be complicated.

It is.

AGENTS.md was supposed to fix this. One Markdown file, no special syntax, every tool reads it. And for the main instructions file, it mostly works. But modern AI coding tools have moved beyond a single file. They support rules, commands, reusable workflows. And every single tool invented its own directory structure for those. None of them interchangeable.

There is no good solution for this. I looked. The best I could come up with is symlinks — filesystem hacks so each tool finds the same content in the one specific directory it’s willing to look at. It works. It’s also absurd.

The real irony: Anthropic co-founded the Agentic AI Foundation that stewards AGENTS.md. Their own tool, Claude Code, doesn’t read it. The feature request has 2,500+ upvotes. No official response.

None of this is technically hard. Reading one extra file, scanning one shared directory, is seemingly trivial. This is a coordination problem. And right now, developers like me are the ones duct-taping over the gaps.