orchestrator

To prevent AI hallucinations, you decompose large specifications by reading an approved spec.md, mapping BDD scenarios to DDD components, and emitting a topologically ordered tasks.md file with atomic tasks under 100 LOC and ~500 lines of context each.

You map BDD Given/When/Then scenarios into DDD tactical components by transforming them into Factories, Repositories, Use Cases, Entities, Value Objects, and Domain Events during the task decomposition process to establish explicit dependencies and topological execution order.

The best way to break down a multi-thousand line spec is to generate a changes/[change-id]/tasks.md file containing Phase 0-4 tasks with strict context limits of ~500 lines, clear acceptance criteria, and explicit dependencies to enable deterministic developer progress.

Yes, you need an approved spec.md before generating atomic implementation tasks. The decomposition process reads the approved spec.md to map BDD scenarios to DDD components and emit a topologically ordered set of tasks ready for phase-by-phase implementation.

Task decomposition limits context to 500 lines per task to eliminate the risk of AI hallucinations, ensuring that each atomic task under 100 LOC is easy to implement, verify, and maintain without being overwhelmed by excessive surrounding specification text.

Yes, you can parallelize implementation work safely because topologically ordered task decomposition establishes explicit dependencies between atomic tasks, enabling safe parallel work and predictable delivery across the generated phases.