An agentic coding tool begins each session with an empty context window. It knows nothing about your project unless you've just told it or it's read it. That blank slate is why the same tool can feel sharp one session and scattered the next. A CLAUDE.md or AGENTS.md file is how you hand it a standing set of instructions so that it stops improvising the ground rules every time. Markdown files are the least glamorous feature in this whole series but one of the highest-leverage tactics to use.
What These Files Are
CLAUDE.md and AGENTS.md are exactly what they look like (if you’ve ever opened one). They are ordinary text files written in plain markdown and placed in your project folder (and optionally in your home directory for personal defaults). The agentic tool reads the file into its context at the very start of every session, before your first request, and everything it generates afterward is conditioned on what the file says.
The filename depends on the tool. Claude Code reads a file named CLAUDE.md. OpenAI’s Codex (and a growing number of other tools) read one named AGENTS.md. These files and reading on tool start have become a shared convention across the AI industry. The idea is identical; only the label differs. Be aware that they do not transfer between tools. CLAUDE.md is Claude Code’s, AGENTS.md is Codex’s, and switching tools means providing the context again. Each tool is its own island. (Pro tip: You can maintain a copy of both in your project folders and, whenever closing out, ask the tool you’re using to make sure they are identical).
What goes inside these files are the things you would re-explain in every session. Things like how the project is laid out, the exact commands to build and test it, the coding conventions you expect, your standing “always do this, never do that” rules, and which directories hold sensitive material. These files can live at several levels at once, from your personal defaults down to a specific subfolder, with the more specific instructions adding to the broader ones.
Behavior analysts have a name for what is happening here. You could shape the tool’s behavior the slow way, correcting it session after session until it finally does things your way. Or you can supply the rule up front, in writing, and its output comes under the control of that rule from the first moment. In our vocabulary, this is closer to rule-governed responding than to contingency-shaped responding. This is technically a structural analogy because agentic tools are software, not humans. But the practical effect is one we would predict. A written rule, present before the behavior, produces more consistent behavior than waiting to correct each error after it happens.
