bmad-project-context

Run the skill with the setup intent on a repository that has no meaningful instruction file. It scans the repo, interviews you about rules no scan can find, verifies every claim, and writes a small approved block between managed markers in AGENTS.md.

Use the adopt intent, which treats your existing instruction file as the baseline. Every instruction enters a ledger with a retain, rewrite, relocate, automate, or delete decision you approve, and the old file is typically reduced to an import line once loading is verified.

Only what changes agent behavior: policy code cannot express, command caveats configs do not state, conventions differing from ecosystem defaults, observed pitfalls, and cross-component rules. Directory trees, stack lists, and style rules enforced by linters are excluded.

No. Deletion requires one of four explicit grounds: stale or incorrect, mechanically enforced by a check, harmful or contradictory, or line-item user approval. Every proposed removal is shown in a ledger before anything is written.

Refresh reads the recorded provenance SHA, re-verifies every path and caveat, and runs git log for deletions and renames since that commit. Lines whose evidence is gone become ledger entries for approval, never silent edits.

Only when the rules are subtree-exclusive, substantial, materially reduce the root block, and loading is verified for every harness in use. Otherwise rules stay at the root as path-qualified lines, which cost less than a file no agent loads.