bmad-project-context

Invoke the skill with the setup intent against your repository. It scans configs, CI, and source for verifiable facts, interviews you only for what scans cannot answer, then shows the complete block for approval before writing it between markers in AGENTS.md.

Include only what agents cannot derive from the repo: policy like frozen paths and branch rules, caveats about running tests, conventions differing from ecosystem defaults, and pitfalls with observed evidence. Exclude directory trees, stack lists, and commands already stated in package.json or CI config.

Use the record intent with the task, the mistake, the correction, and its evidence. One occurrence is noted; a recurring or costly mistake earns a pitfall line or an exact command invocation in the block, and mechanically preventable issues are proposed as hooks or CI checks instead.

No. It splices content only between the bmad:context markers, leaving everything outside byte-identical, and never commits. When other instruction files contradict the block, it proposes fixes but requires your approval for every write.

Refresh reads the provenance line's commit SHA, re-verifies every path and caveat, and runs git log for deletions and renames since that SHA against every line. Lines whose evidence is gone are updated or removed, and the interview shrinks to what changed.

Rules bounded to a specific directory go in a nested AGENTS.md at that location, attached by path rather than by pointer. A component or child repo gets its own file only when work keeps landing there and its truths do not belong at the parent level.