Subagent Driven Development for Claude Code

Use Claude Code's task tools and the Agent tool to coordinate implementation and review. The orchestrator owns readiness, dependency decisions, and the correction loop. Subagents investigate, implement, and review the work assigned in their prompts.

Supported tools

This skill uses only these Claude Code tools:

• TaskCreate accepts { subject, description, activeForm?, metadata? } and creates a task with status pending.
• TaskUpdate accepts { taskId, status?, subject?, description?, activeForm?, owner?, metadata?, addBlockedBy?, addBlocks? }. Its status values are pending, in_progress, and completed; deleted permanently removes the task. addBlockedBy records the task IDs that must complete before this task can start.
• TaskList accepts {} and returns each task's id, subject, status, owner, and open blockedBy IDs.
• TaskGet accepts { taskId } and returns one task's full subject, description, status, blocks, and blockedBy.
• Agent accepts { description, prompt, subagent_type, run_in_background? } and runs one subagent. Several Agent calls in one message run in parallel. The required subagent_type values are worker and reviewer; scout and oracle are optional. When installed through the plugin, these agents are namespaced, such as agentic-workflow:worker; use whichever form is listed.

The task tools are not enabled by default on every model. If TaskCreate is unavailable, stop and tell the user to enable the task tools, for example by setting CLAUDE_CODE_ENABLE_TODO_TOOLS=1; do not substitute another tracker. If the task tools are listed but not loaded, load them before use. If the worker or reviewer subagent is unavailable, stop and report it; do not substitute another agent type.

Keep task objects to the fields supplied by these tools. Record the role in metadata.role, dependencies with addBlockedBy, and scope, acceptance, and evidence in description; do not invent a scheduler.

Invariants