parallelize-batch-work
Parallelize Batch Work — Fan Out, Don’t Sequence
Section titled “Parallelize Batch Work — Fan Out, Don’t Sequence”Trigger: ≥2 independent units, not a size threshold. Under-spawning costs more than over-spawning (an over-spawn is only reported, enforced=false). Independence check: share no mutable state, no sequential dependency, no design-intent, else DAG or coupled; the token-multiplier gate covers AI-subagent fan-out only. Count units before sequencing: default = parallelize, sequencing needs a stated reason; Never write a sequential loop over a batch.
Write-capable fan-out: the orchestrator decides isolation up front, per lane — git worktree add when lanes may touch one file, need separate branches/PRs, rebase or concurrent full suites; otherwise don’t. Different repos are always safe; when you fan out, say the isolation choice once.
Never idle on a gate the next unit does not read — name the dependency or start now; prepare draft-PR work that fails safe.
Batch, verify once. Never the edit→verify→wait→fix loop one change at a time on a large task. Availability: delegate-to-preserve-context.md. Full text: docs/parallelize-batch-work.md · docs/ai-velocity-batch-compile.md.