Give your AI standing instructions with an AGENTS.md file

I went live to set up standing instructions for an AI agent with an AGENTS.md file, and I started out assuming my global CLAUDE.md could just be renamed. It turned out Claude Code only reads AGENTS.md at the project level, so I changed course, split my instructions up, and put a project-level file in an R project to see what loaded.

You'll learn:

• How to put instructions you keep repeating, like which package to use, into one file your agent reads every time instead of restating them in each prompt
• The difference between a global CLAUDE.md and a project-level AGENTS.md, including that Claude Code reads AGENTS.md per project but needs CLAUDE.md for the global file
• A way to split a long global instruction file into general rules, path-specific rules that load by file type, and skills, so the agent does not load everything every time
• A way to check that your instructions are working: ask the agent which files it loaded, then look at its output for the style and file-saving behavior you specified

Tools mentioned: Claude, Claude Code, Solo, ally, Positron, tidyverse, Laravel, Rough Draft, Dia

Key moments:

• 00:14 Introducing standing instructions with AGENTS.md
• 01:51 Matt Pocock's guide to AGENTS.md
• 03:56 Asking Claude whether it reads AGENTS.md
• 05:14 Global file must still be CLAUDE.md
• 07:24 The ally function for editing agent instructions
• 08:02 Walking through my global CLAUDE.md
• 10:42 Claude's suggestion for splitting up the file
• 12:04 Skills versus instructions in the file
• 15:59 What rules are and how they load
• 17:38 Creating a project-level AGENTS.md
• 20:12 Example AGENTS.md from an R package project
• 26:31 Testing with a request to describe the visits data
• 26:52 Confirming both global and project files load
• 28:52 Checking output follows the saved-output instructions