subagent-driven-development

Use when executing implementation plans with independent tasks in the current session

Install with Tessl CLI

npx tessl i github:obra/superpowers --skill subagent-driven-development

What are skills?

1.27x

Quality

46%

Does it follow best practices?

Impact

97%

1.27x

Average score across 8 eval scenarios

Optimize this skill with Tessl

npx tessl skill review --optimize ./skills/subagent-driven-development/SKILL.md

SKILL.md

Review

Evals

Evaluation results

100%

55%

Execute a Three-Task Implementation Plan

Orchestrating a multi-task implementation plan with correct review ordering and task tracking

Criteria

Without context

With context

Extract all tasks and create TodoWrite upfront

25%

100%

Fresh subagent per task

100%

Spec compliance review before code quality review

100%

Final code reviewer after all tasks

80%

100%

Mark tasks complete in TodoWrite

26%

100%

Without context: $0.2749 · 1m 37s · 14 turns · 130 in / 5,122 out tokens

With context: $0.5634 · 2m 38s · 22 turns · 758 in / 8,274 out tokens

100%

47%

Write Subagent Dispatch Prompts

Constructing spec reviewer and implementer subagent prompts with correct patterns

Criteria

Without context

With context

Full task text in implementer prompt

100%

Context or scene-setting in implementer prompt

100%

Spec reviewer distrust instruction

33%

100%

Spec reviewer checks for missing AND extra work

100%

Implementer asked to self-review before reporting

30%

100%

Without context: $0.2805 · 1m 20s · 19 turns · 25 in / 3,819 out tokens

With context: $0.3211 · 1m 27s · 13 turns · 793 in / 4,170 out tokens

100%

35%

Execute a Five-Task Plan with Independent Tasks

Sequential task execution with no parallel implementation subagents

Criteria

Without context

With context

One implementer at a time

90%

100%

Full review cycle before next task

26%

100%

Fresh subagent for each task

100%

Final review after all tasks

50%

100%

Without context: $0.2191 · 1m 16s · 11 turns · 18 in / 3,702 out tokens

With context: $0.2913 · 1m 21s · 14 turns · 19 in / 3,830 out tokens

85%

-5%

Dispatch Code Quality Reviewer for a Completed Task

Constructing code quality reviewer dispatch with correct prerequisites and inputs

Criteria

Without context

With context

Spec compliance passes before code quality dispatch

66%

50%

Git SHA references included

100%

Implementation description and requirements included

100%

Fix loop for quality issues

100%

Without context: $0.1771 · 1m 6s · 11 turns · 16 in / 2,644 out tokens

With context: $0.3169 · 1m 24s · 15 turns · 17 in / 3,720 out tokens

100%

Handle Review Failures in Subagent Workflow

Handling review failures with correct fix-and-re-review loop

Criteria

Without context

With context

Implementer fixes spec reviewer issues

100%

Spec re-review after fix

100%

Code quality review only after spec passes

100%

Task not marked complete with open issues

100%

Without context: $0.1395 · 52s · 9 turns · 14 in / 2,634 out tokens

With context: $0.4452 · 2m · 20 turns · 2,000 in / 5,652 out tokens

94%

26%

Plan a Subagent Dispatch Strategy

Model selection by task complexity

Criteria

Without context

With context

Cheap model for Task 1

100%

Standard model for Task 2

100%

Most capable for Task 3

100%

Cheap model for Task 4

100%

Most capable for spec reviewer

100%

Most capable for code reviewer

100%

Rationale references file scope

100%

Rationale references spec completeness

100%

Cost/speed framing

100%

50%

Without context: $0.1182 · 51s · 8 turns · 13 in / 2,131 out tokens

With context: $0.2200 · 1m 7s · 12 turns · 15 in / 2,773 out tokens

100%

Respond to Implementer Reports

Handling implementer status escalations

Criteria

Without context

With context

DONE_WITH_CONCERNS: reads concerns

100%

DONE_WITH_CONCERNS: correctness addressed

80%

100%

DONE_WITH_CONCERNS: file-size is observation

100%

NEEDS_CONTEXT: provides answer

100%

NEEDS_CONTEXT: re-dispatches implementer

100%

BLOCKED: assesses blocker type

100%

BLOCKED: task decomposition or escalation

100%

BLOCKED: no same-model retry

100%

No manual fixes

100%

No ignoring escalation

100%

Without context: $0.1691 · 59s · 9 turns · 14 in / 2,157 out tokens

With context: $0.2584 · 1m 17s · 13 turns · 252 in / 3,276 out tokens

100%

Start a Feature Development Session

Branch isolation setup and answering implementer questions

Criteria

Without context

With context

Uses git worktrees

100%

Not starting on main/master

100%

Setup before dispatch

100%

All questions answered

100%

Answers before implementation

100%

Clear and complete answers

100%

Re-dispatch after answers

100%

TodoWrite setup

100%

Plan read upfront

100%

Without context: $0.2730 · 1m 4s · 11 turns · 14 in / 2,915 out tokens

With context: $0.3167 · 1m 26s · 16 turns · 256 in / 4,053 out tokens

Evaluated: 1 day ago
Agent: Claude Code
Model: Claude Sonnet 4.6

Table of Contents

Execute a Three-Task Implementation Plan Write Subagent Dispatch Prompts Execute a Five-Task Plan with Independent Tasks Dispatch Code Quality Reviewer for a Completed Task Handle Review Failures in Subagent Workflow Plan a Subagent Dispatch Strategy Respond to Implementer Reports Start a Feature Development Session

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.