Superpowers is an open-source agentic skills framework and software-development methodology released under the MIT license that aims to make coding agents like Claude Code, Cursor, GitHub Copilot CLI, and Gemini CLI behave like disciplined engineers rather than fast typewriters. The project has exploded on GitHub with 191,000+ stars and 17,000+ forks since its public launch, becoming one of the most-starred AI engineering frameworks of the year. Rather than another autonomous code generator, Superpowers ships a curated library of composable skills, mandatory workflow gates, and a brainstorm-design-plan-implement loop that the agent must follow before any code touches disk. ## Why Superpowers Matters Most coding agents available today jump straight from a one-line prompt to a wall of generated code, producing solutions that drift from real requirements, mix concerns, and accumulate architectural debt. Superpowers attacks the root cause: the absence of structured engineering process. By making brainstorming, design approval, planning, and review explicit steps that the agent cannot skip, it converts AI-assisted coding from a guessing game into a repeatable methodology. The result is an agent that asks clarifying questions, presents a design for human approval, writes a plan, then executes through subagents with dual-stage review, mirroring how senior engineers actually ship features. ## Composable Skills Library The framework is organized around skills, small reusable workflows stored as markdown that the agent automatically activates when relevant. The core library covers Testing skills for RED-GREEN-REFACTOR test-driven development, Debugging skills for systematic root-cause analysis with verification steps, Collaboration skills for brainstorming, planning, code review, and parallel branch management, and Meta skills that teach the agent how to author new skills and understand the framework itself. Skills compose: a feature implementation may pull in a test skill, a code review skill, and a debug skill in sequence without manual orchestration. ## Mandatory Workflow Gates Unlike documentation-style prompt libraries, Superpowers enforces workflow as code. Before writing implementation the agent runs a brainstorm phase to expose ambiguity, then surfaces a design document for human approval, then produces a detailed implementation plan that subagents execute in parallel. Each subagent's output is reviewed twice, once by an automated check and once by a peer-style review pass, catching the common AI failure modes of partial implementations, missing edge cases, and silent assumptions. The gates make autonomous multi-hour sessions practical without runaway drift. ## Test-Driven and Verification-First Philosophy The authors explicitly favor test-driven development, systematic process over guesswork, simplicity over cleverness, and verification over assumption. Skills like the TDD skill require the agent to write a failing test before any implementation, run the test to confirm it fails for the right reason, write the minimum code to pass, then refactor. Debug skills require the agent to reproduce a bug, form a hypothesis, design a verification experiment, and only then propose a fix. The discipline is unusual for AI agents and visibly raises the floor on output quality. ## Multi-Agent and Platform Support Superpowers is platform-neutral. It integrates with Claude Code, GitHub Copilot CLI, Cursor, Gemini CLI, and any agent that can read markdown skill files and follow structured workflows. The same skill library can be shared across a team's tools, so the brainstorming workflow used in Claude Code on one developer's machine matches the workflow used in Cursor on another's, producing consistent code review and test discipline across the org. ## Limitations The enforced workflow trades speed for discipline. Quick prototypes and throwaway scripts feel heavyweight under mandatory brainstorm-design-plan gates, and developers who prefer iterating directly in code may chafe at the upfront approval steps. The framework is also still rapidly evolving, with skills being added, renamed, and reorganized week to week, so teams pinning to a release will want to track upstream changes carefully. Finally, Superpowers raises the floor on agent output but cannot replace human judgment on architecture, security trade-offs, and product decisions.