bmad-project-context

Run the setup intent to scan the repository, interview for rules no scan can find, and write a small verified block into AGENTS.md. Every claim is path-checked against the repo and you approve the complete block before anything is written.

Use the adopt intent, which treats the existing file as the baseline and opens a ledger entry for every instruction. Each rule is retained, rewritten, relocated, automated, or deleted with a stated reason, and the file typically reduces to an @AGENTS.md import once loading is verified.

Include policy the code cannot express, command caveats configs cannot state, conventions differing from ecosystem defaults, and pitfalls with observed evidence. Exclude directory trees, stack lists, style rules a linter enforces, and anything derivable fresh from the repo.

Yes, a subdirectory can get its own AGENTS.md when its rules are subtree-exclusive, substantial, and loading is verified for every harness in use. Otherwise rules stay in the root block as path-qualified lines, which cost less than a file no session loads.

Deletion requires one of four grounds: the instruction is stale or incorrect, a hook or CI check mechanically enforces it, it is harmful or contradictory, or the user approves the deletion as a line item. Brevity or discoverability alone is never grounds.