This page documents the subagent-driven-development (SDD) skill, a structured workflow for executing implementation plans by dispatching fresh subagents per task with a mandatory two-stage review protocol. This skill is designed for use in environments that support subagent dispatch (Claude Code, Codex, OpenCode, Gemini CLI).
For creating the implementation plans that this skill executes, see writing-plans. For executing plans in environments without subagent support, see Executing Plans in Batches.
Sources: skills/subagent-driven-development/SKILL.md1-9
The SDD workflow is governed by a specific quality formula:
Fresh subagent per task + Task review (Spec Compliance + Code Quality) + Broad final review = High Quality, Fast Iteration
By precisely crafting instructions and context for specialized agents, the system ensures they stay focused. Subagents should never inherit the main session's history; instead, the controller constructs exactly what they need, preserving the controller's own context for coordination work. skills/subagent-driven-development/SKILL.md10-12
Continuous execution: The controller is instructed to execute all tasks from the plan without stopping to check in with the human partner unless a BLOCKED status occurs, ambiguity genuinely prevents progress, or all tasks are complete. skills/subagent-driven-development/SKILL.md17-18
Sources: skills/subagent-driven-development/SKILL.md10-18
The system uses a decision matrix to determine if SDD is appropriate for the current task.
Comparison with executing-plans:
| Feature | subagent-driven-development | executing-plans |
|---|---|---|
| Context | Fresh subagent per task (no pollution) | Single session history |
| Review | Task-scoped (Spec + Quality) | Single review or batch review |
| Speed | Faster iteration (no human-in-loop) | Manual handoffs/checkpoints |
| Session | Remains in current session | Often involves parallel sessions |
Sources: skills/subagent-driven-development/SKILL.md19-44
The SDD process follows a strict loop for every task defined in the implementation plan. Recent optimizations have merged spec and quality reviews into a single subagent dispatch to reduce overhead. docs/superpowers/specs/2026-06-09-sdd-task-scoped-review-dispatch-design.md54-59
Sources: skills/subagent-driven-development/SKILL.md45-83 docs/superpowers/specs/2026-06-09-sdd-task-scoped-review-dispatch-design.md54-59
To optimize for cost and speed, the system selects models based on the mechanical vs. creative nature of the task. skills/subagent-driven-development/SKILL.md99-101
Crucial: Turn count beats token price. Cheap models often take 2-3× the turns, increasing wall-clock time and total cost. A mid-tier model is recommended as the floor for reviewers and prose-based implementers. skills/subagent-driven-development/SKILL.md119-122
Sources: skills/subagent-driven-development/SKILL.md99-126
Implementer subagents report their status using one of four specific tokens. The controller (main agent) responds based on the protocol:
| Status | Meaning | Controller Action |
|---|---|---|
| DONE | Work finished confidently. | Proceed to Task Review. skills/subagent-driven-development/implementer-prompt.md127 |
| DONE_WITH_CONCERNS | Finished, but flagged doubts (e.g., file size). | Read concerns. Address correctness/scope before review. skills/subagent-driven-development/implementer-prompt.md136 |
| NEEDS_CONTEXT | Information missing. | Provide missing context and re-dispatch. skills/subagent-driven-development/implementer-prompt.md137 |
| BLOCKED | Cannot complete task. | Assess: Provide context, upgrade model, or break task into smaller pieces. skills/subagent-driven-development/implementer-prompt.md137-138 |
Sources: skills/subagent-driven-development/implementer-prompt.md125-138
Dispatched via implementer-prompt.md. skills/subagent-driven-development/implementer-prompt.md1-139
Dispatched via task-reviewer-prompt.md. skills/subagent-driven-development/task-reviewer-prompt.md1-166
Sources: skills/subagent-driven-development/implementer-prompt.md1-139 skills/subagent-driven-development/task-reviewer-prompt.md1-166
The SDD skill bridges Natural Language requirements (the Plan) to Code Entities (the Implementation) through structured tool calls and temporary files.
Key Data Flow Elements:
scripts/task-brief PLAN N to provide high-fidelity context without polluting the controller's context. docs/superpowers/specs/2026-06-09-sdd-task-scoped-review-dispatch-design.md89-91git diff to a temporary file (/tmp/sdd-task-N.diff) which the reviewer reads once, avoiding massive text pastes into the prompt. docs/superpowers/specs/2026-06-09-sdd-task-scoped-review-dispatch-design.md64-68<git-dir>/sdd/progress.md to track state across session resumes. docs/superpowers/specs/2026-06-09-sdd-task-scoped-review-dispatch-design.md91-92Sources: docs/superpowers/specs/2026-06-09-sdd-task-scoped-review-dispatch-design.md54-92 skills/subagent-driven-development/task-reviewer-prompt.md32-44
SDD is the execution phase of a larger pipeline and relies on several other skills:
Sources: skills/subagent-driven-development/SKILL.md63-66 docs/superpowers/specs/2026-06-09-sdd-task-scoped-review-dispatch-design.md1-5
Refresh this wiki