Skip to main content
Helix reads your project’s AGENTS.md or CLAUDE.md and places it last in the prompt — the final thing the model reads before your turn, framed as your project’s working rules. Put your build commands, conventions and house rules there.

​
Which file wins

Discovery walks from the working directory up through every ancestor. In each directory the first of these names wins and the rest are skipped — one file per directory, never merged: AGENTS.override.md → AGENTS.md → AGENTS.MD → CLAUDE.md → CLAUDE.MD The collected files render root-first with the nearest file last, so the closer a file is to where you launched, the closer it sits to the model — and the more weight it carries when files disagree. With this tree, launched from packages/api/:
Both apply; packages/api/AGENTS.md has the final word. An AGENTS.md and a CLAUDE.md in the same directory do not combine — AGENTS.md is earlier in the candidate order, so only it loads.

​
Placement in the prompt

Your files close the assembled prompt — the last block before your turn, announced to the agent as project instructions to follow. That placement is deliberate, and it is why the file actually gets obeyed. Simplified to its skeleton (real block names, stand-in bodies), the prompt the model receives looks like this:
Each of your files arrives verbatim in its own <project_instructions> block, labelled with its path, root first and nearest last. It loads in every mode — the orchestrator, agent mode, and Prime — and it reaches the sub-agents Helix dispatches, which matters more than it sounds: an agent editing files in your repository without your conventions will violate every one of them. One exclusion, also deliberate: the binary ships internal files in its own bundle directory, and those are never rendered as your project context — only files you wrote reach the block.