subagent-driven-development

Subagent-Driven Development is a specialized architectural skill within the Superpowers methodology, designed to optimize autonomous software engineering. It focuses on the principle that complex implementation plans should be broken down into discrete, independent tasks, each handled by a fresh, context-isolated subagent. By avoiding the accumulation of stale session context, this skill ensures that every subagent operates with a clean state, focused strictly on its specific assignment, which significantly reduces errors and hallucinations during the coding process. This skill is primarily intended for developers and coding agents operating within the Superpowers framework who require high-velocity, high-fidelity execution of pre-defined technical roadmaps.

• Employs a mandatory two-stage review protocol after every task: first validating adherence to the technical specification, then evaluating overall code quality and maintainability.

• Ensures complete isolation by constructing task-specific prompts, preventing context pollution and maintaining a narrow scope of work for each specialized agent.

• Supports intelligent model selection based on task complexity; assigns mechanical tasks to cost-efficient models while reserving high-capability models for complex architectural or integration challenges.

• Automates the workflow from task dispatch to verification, utilizing TodoWrite tracking to manage progress across multi-task implementation plans.

• Seamlessly integrates with TDD (Test-Driven Development) and Git worktrees to maintain a stable, test-verified codebase throughout the development lifecycle.

• Use this skill when you have a well-defined, multi-task implementation plan where tasks exhibit low coupling and can be executed within the current environment.

• Inputs typically include an approved implementation plan, clear task definitions, and established technical constraints.

• Outputs consist of verified, committed code changes that have passed both automated test suites and double-layered manual/agentic reviews.

• Avoid using this skill for highly coupled, monolithic tasks where inter-dependency requires persistent, shared context across the entire session.

• Requires consistent use of TodoWrite to track completion status and handle subagent statuses like DONE or DONE_WITH_CONCERNS.