Superpowers is an open-source collection of composable skills and workflow methodology for coding agents. Rather than letting an agent improvise from a blank slate, it enforces a structured development cycle: requirements clarification, time-boxed planning with exact file paths, subagent-delegated implementation, TDD cycles, and two-stage code review — all wired together as slash commands.
What It Is
Superpowers is a set of structured prompts and slash-command workflows (.claude/skills/ compatible) designed to make coding agents follow a disciplined software development process. It works with Claude Code, Cursor, GitHub Copilot CLI, Gemini CLI, OpenAI Codex, and OpenCode.
The core insight: coding agents produce worse output when they jump straight to implementation. Superpowers forces the agent to first clarify requirements, then produce a plan (targeting 2–5 minutes per task) with explicit file paths and change types, then delegate subtasks to subagents, and finally run two review passes: spec compliance first, code quality second.
Why It Belongs in Trial
- Active release cadence (v5.0.7, March 2026), MIT-licensed, with 28+ contributors
- The two-stage review pattern (spec → quality) addresses a real failure mode in agentic coding: agents that produce technically correct code that misses the stated requirement
- The methodology operationalizes the "plan before code" discipline that separates reliable agentic workflows from brittle one-shot prompting — giving teams a reference implementation to evaluate
- However, it is opinionated about process in ways that may not suit all teams or workflows; hands-on evaluation is warranted before broad adoption
Core Workflow Stages
| Stage | What Happens |
|---|---|
| Clarify | Agent asks targeted questions before starting — surfaces ambiguity upfront |
| Plan | Produces a structured plan (file, change type, rationale) targeting 2–5 min per subtask |
| Implement | Delegates to subagents with scoped context; uses TDD (RED → GREEN → REFACTOR) |
| Review 1 | Checks: does the implementation satisfy the stated specification? |
| Review 2 | Checks: is the code quality, maintainability, and style acceptable? |
Design Principles
- Simplicity is the primary goal — agents default to the simplest solution that satisfies requirements
- Evidence-based completion — "done" means tests pass and behavior is verified, not "I think it's right"
- Subagent isolation — complex tasks are decomposed; each subagent gets focused, clean context
Relationship to Harness Engineering
Superpowers is a skills-layer contribution to the harness engineering stack. It doesn't replace a CLAUDE.md or meta-harness; it provides opinionated workflow templates that sit above the harness, guiding the agent through the development lifecycle rather than just configuring its environment.
GitHub: github.com/obra/superpowers