Instructions for agents
On this page
An agent that starts every session from zero repeats your explanations. A short instructions file at the project root, committed with the code, gives every agent and every session the same starting knowledge.
One file, read by all three
Codex and the Antigravity CLI read AGENTS.md at the project root. Claude
Code reads CLAUDE.md. Keep one source of truth: write AGENTS.md, and make
CLAUDE.md a one-line pointer to it.
Read your instructions from AGENTS.md; this project keeps nothing in
agent-specific files.
What belongs in it
- What the project is, in two sentences, and where the longer brief lives.
- How to set up, build and test, as the exact commands.
- What must not be done without asking: the files that are generated, the branches that are protected, the data that is real.
- How you want work reported: what a finished change includes, what a commit message says.
- Where decisions and status are recorded, so the agent reads them before acting and writes to them when done.
Keep it short and current; an instructions file that lies is worse than none. DevCapsule's own repository uses this pattern with a workflow that records status, decisions and next steps in files, so that any agent, or any person, can resume cold; Collaboration describes it, and you can adopt it or not.