Batch

Invoke as /afk:batch. Use it when a plan splits into 5–30 units that are each independently mergeable as their own PR: a codebase-wide migration, rename, dependency bump, or the same pattern change repeated across many files.

batch decomposes the plan into non-overlapping units, then runs one worktree worker per unit in parallel — each committing, pushing, and opening its own PR.

What it does

Batch is the parallel, PR-per-unit alternative to afk:implement. It decomposes the work into non-overlapping independent units, presents the decomposition for user approval, then spawns one background worker per unit in its own git worktree. Each worker implements, runs tests, optionally runs afk:simplify on substantial diffs, commits, pushes, and opens a PR, reporting back a single PR: <url> line.

• Units must be independently mergeable (no shared interface still being designed) and non-overlapping (no two units touch the same file).
• If a clean independent split is not possible, Batch stops and routes to the afk:implement freeze-then-fan-out recipe instead.
• Workers run concurrently; the lead tracks progress in a live status table and triages any failures.

Output artifact: a status table of all units, and a final summary: Batch: N/total units as PRs, list of PR URLs, and any failed/blocked units.

View the full skill on GitHub