asermax Load mahou:basics first. Then read .mahou/human-review.md if present.
Human review
Verifies the user's own reading of a change and settles it with them. Nothing is edited until they rule on it.
When both review skills run, mahou:agentic-review goes first. It clears the mechanical findings so this pass spends attention on judgment instead.
Step 1: Establish the changeset
Load mahou:changeset. It gives you the repository paths you dispatch against, and the rule that tells you whether what an annotation flags was introduced by this change or predates it.
Step 2: Collect the annotations
Start one plannotator review per repository, backgrounded, all in a single message:
cd {absolute-repo-path} && plannotator review
Backgrounding is what keeps that cd safe, because it never changes the session's working directory. Address repositories everywhere else the way changeset says.
Keep a map of repository to task so every result is attributed. Never poll. Completions arrive on their own.
Step 3: Work each repository as its annotations land
Handle a repository the moment its result arrives, start to finish, and never hold a finished one behind a slow one.
Verify
Annotations are unverified input. They carry what the user wants, not always the full context.
Read the flagged lines before responding. Establish whether the change introduced what is flagged or whether it predates the change, which changeset covers. Then give each annotation a verdict, either Confirmed, Partly, Not a bug, or Intended, with the lines that settle it quoted.
Report the base plannotator used alongside the changeset base. They can differ, and when they do the annotations cover a smaller diff than the changeset does.
Present
Not every annotation is a finding. Some approve, some remove, some ask a question. Answer the questions first, because an answer changes what gets applied.
Change nothing yet. Never apply the obvious ones unasked. Never argue a finding down: if you believe it is wrong, say why in one line and let the user rule. A "Not a bug" verdict is your reading, not a decision. The user routinely overrides it, and that is the correct outcome.
Apply what the user approves
Before writing code, read the project's CLAUDE.md and .mahou/basics.md when present, and read the wiki entries for the technologies the fix touches. Follow the conventions of the repository you are in.
Run that repository's linters and tests afterwards, scoped to the files you touched rather than the whole tree. A blanket formatter run reformats unrelated files and drags them into the change.
Leave everything uncommitted.