skill-subagent-driven-development

Read the approved plan once, extract each task with the task-brief script, and dispatch a fresh implementer subagent per task. After each task, generate a scoped diff with review-package and dispatch a task reviewer before marking the task complete.

Subagent-driven development gives every task a fresh implementer and reviewer context with automatic review checkpoints in one coordinating session. Direct plan execution suits tightly coupled tasks, handoffs to another session, or work without commit authorization.

Use parallel dispatch only when independent lanes can run concurrently with disjoint write ownership. Inside this skill, implementers always run sequentially in one workspace to avoid conflicts.

Assess the blocker: provide missing context and re-dispatch, escalate to a more capable profile for harder reasoning, split oversized tasks, or reconcile plan drift. Never force the same model to retry without changes.

Pasted content stays resident in the controller's context and is re-read every turn, bloating later dispatches. File handoffs via task-brief, report files, and review packages keep each subagent's context limited to its own task.

It costs more subagent invocations and controller preparation than direct execution, and it requires tasks separable enough for isolated briefs. It must not be used as the top-level topology while plan-bound conditions of satisfaction are active.