Instruction files exist because every new agent session starts without context. Claude Code reads CLAUDE.md files from the project and the user's settings. AGENTS.md is an open format meant as one predictable place for coding-agent instructions, and its site lists support in tools such as OpenAI Codex, Cursor, GitHub Copilot, Jules and many others. The content is similar everywhere: how to build and test, code style, where things live, what never to do.
They work well for stable, short facts. The trouble starts when they become the place where the process lives. Each incident adds a rule, each rule adds an exception, and the file grows into a rulebook the agent reads in full and follows in part. Thoughtworks named the pattern "agent instruction bloat" on its Technology Radar. A long file costs context on every session, mixes unrelated rules together, and gives no way to tell which rule applied to a given change.
A practical split keeps the file short. Facts that every session needs stay in it: what the project is, how to run it, where standards live. Procedures move out into something the agent can be handed one step at a time. Rules move into standards documents that a step points to when they apply. Lessons move into memory that is recalled when relevant.