I can never get my AGENTS.MD to be reliably followed.
A good example is the useMemo or useCallback. It frequently loves to add those anyway on more vibe-codey parts of my code where I’m experimenting unless I have a very solid architectural pattern defined.
more context used = less reliable to follow rules
too long rules = bad
too many tools/mcp servers = bad
Thanks for the advice - I’ve noticed I’m on about 1000 lines now for AGENTS.MD which I generally let it do its own thing so will give that a trim.
I don’t have any MCP tools though and not convinced on the context as it happens early into a chat
1,000 lines for an AGENTS.md is a lot! You can read this blog post—a typical AGENTS.md is only about 300 lines. Are you working on a monorepo? Why is that file so long?
If you need a code guide, it is better to use ESLint and add these rules:
Then, in your instructions, ask the AI to always run “lint --fix” after finishing a task. If a warning or error shows up, the model will try to fix it. Don’t add unnecessary code styles to agents; use a linter for that. Or add rules and apply them intelligently or manually.
All good suggestions thanks.
I trimmed my AGENTS.MD to circa 99 lines.
Using Composer-1 I still get this crap to support made-up backwards compatibility despite line “Never support backwards compatibility.
It still peppers XML everywhere despite line 14 saying this.
“Fake news” is how I sum up these stupid files
