Superpowers

What It Does

Superpowers is a software development methodology packaged as composable Agent Skills markdown files that AI coding agents read and follow during development. Created by Jesse Vincent (Prime Radiant, formerly Perl project lead and Keyboardio co-founder) and launched October 2025, it enforces a seven-phase workflow: socratic brainstorming, git worktree isolation, micro-task planning, subagent-driven implementation, TDD (RED-GREEN-REFACTOR), structured code review, and branch completion.

The core mechanism is instruction-following enforcement rather than runtime control: skills are markdown files with YAML frontmatter that agents load via the Agent Skills Specification standard. Agents that support the spec (Claude Code, Codex, Cursor, Gemini CLI, OpenCode, GitHub Copilot CLI) auto-discover skills from the project’s skill directory. The framework’s “enforcement” of TDD and workflow phases relies on the agent actually reading and following those instructions — which differs meaningfully from a hard runtime constraint.

Key Features

• Seven-phase enforced workflow: Brainstorm → Git worktree setup → Plan (2-5 minute micro-tasks with exact file paths) → Subagent-driven execution → TDD (RED-GREEN-REFACTOR mandatory) → Two-stage code review (spec compliance then quality) → Branch completion.
• TDD deletion rule: Agents are instructed to delete any code written before its corresponding test; enforced by skill instruction rather than runtime guard.
• Subagent dispatch: Fresh subagents handle individual tasks, with each subagent receiving isolated context — reducing context contamination between tasks.
• Git worktree isolation: Each development branch uses a dedicated git worktree so the main branch is never directly touched during implementation.
• Cross-platform compatibility: Distinct integration configurations for Claude Code (Claude plugin marketplace), Codex CLI, Cursor, Gemini CLI (gemini-extension.json), OpenCode, and GitHub Copilot CLI; cross-platform test suite covers all six.
• Brainstorm server (v5.0+): Node.js zero-dependency HTTP server generates HTML visual mockups in-browser during the brainstorm phase for design-heavy features.
• Composable skill library: 14 named skills including brainstorming, writing-plans, test-driven-development, systematic-debugging, dispatching-parallel-agents, using-git-worktrees, requesting-code-review, writing-skills.
• Inline self-review (v5.0): Replaced the original dedicated review subagent with inline self-review, reducing review time from ~25 minutes to ~30 seconds (creator’s measurement; no independent verification).
• Active versioning: v5.0.7 as of March 31, 2026; architectural changes tracked in detailed release notes.

Use Cases

• Production feature development: Multi-session builds of well-scoped features where output quality and test coverage are priorities and workflow overhead is acceptable.
• Team coding standards enforcement: Teams can add custom skills to the library encoding their own coding standards, review checklists, and deployment procedures — agents inherit the standards without per-engineer configuration.
• Onboarding AI agents to a codebase: The brainstorm and planning phases force requirement clarification before any code is generated, reducing wasted implementation cycles on misunderstood specs.
• Learning structured agentic development: The explicit seven-phase workflow makes the agent’s development process transparent and reviewable for developers learning how to work effectively with coding agents.

Adoption Level Analysis

Small teams (<20 engineers): Fits well for production-quality feature development. Zero infrastructure overhead — the framework is just markdown files in your repo. The workflow phases enforce code quality discipline that small teams may otherwise skip under deadline pressure. Not appropriate for exploratory prototyping, bug fixes, or one-off scripts where the overhead exceeds the benefit.

Medium orgs (20–200 engineers): Good fit with the caveat that adoption requires buy-in on the methodology. The framework is most valuable when all developers on a codebase use it consistently — inconsistent adoption creates confusion about when the workflow applies. The ability to add team-specific skills (coding standards, deployment checklists) makes it a reasonable investment for teams already standardizing on Agent Skills-compatible tools.

Enterprise (200+ engineers): The framework itself is lightweight enough for enterprise use, but the methodology is opinionated and its enforcement depends on agent instruction-following rather than access controls. Large engineering organizations with heterogeneous tooling across teams will find cross-platform consistency harder to maintain. No enterprise-specific features (RBAC, audit trails, centralized skill management).

Alternatives

Evidence & Sources

• GitHub repository (obra/superpowers) — primary source; README, skills, release notes, cross-platform integrations
• Superpowers: How I’m using coding agents in October 2025 (creator’s blog) — most honest account of what was shipped vs. incomplete at launch
• The Superpowers Plugin for Claude Code (builder.io) — substantive workflow analysis with explicit limitations (environment debugging, spec inheritance errors)
• Superpowers Skills Framework hits 121k stars (byteiota) — growth metrics and community adoption data
• Stop AI Agents from Writing Spaghetti: Enforcing TDD with Superpowers (yuv.ai) — TDD enforcement mechanism analysis
• Superpowers Framework on Star History (star-history.com) — independent star growth tracking

Notes & Caveats

• Enforcement is instruction-following, not runtime control. The TDD deletion rule, code review gates, and workflow phases are implemented as agent instructions in SKILL.md files. Agents can and do deviate from instructions when context is complex, instructions conflict, or the model’s judgment overrides them. This is fundamentally different from a test runner that blocks deployment on coverage failure.
• v5.0 removed the independent review agent. The original two-stage review (a fresh subagent with no implementation context reviewing the implementer’s work) was the framework’s most defensible quality mechanism. v5.0’s inline self-review is faster but loses the independent perspective. Teams that valued the original review architecture should assess whether v5.x meets their needs.
• Star count outpaces production evidence. 151k+ GitHub stars is exceptional; the framework is clearly resonant with the developer community. However, star counts for methodology frameworks skew high relative to actual deployment: developers star promising approaches at research stage. Independent production case studies beyond anecdotal community reports have not been published.
• Scope limitations are explicit. The creator documents that Superpowers is not appropriate for environment debugging, quick bug fixes, exploratory prototyping, or single-file scripts. Teams should evaluate against their actual development distribution — if 60% of daily work is bug fixes and exploratory tasks, the benefit window is narrower than the marketing suggests.
• No independent benchmarks. The 85–95% test coverage claim circulating in community channels is self-reported by users; no controlled study compares Superpowers-guided agents against baseline agents on identical task corpora. The absence of benchmarks was noted as a criticism by GitHub commenters.
• Active maintenance but small contributor base. 31 contributors as of April 2026 with Jesse Vincent as the dominant committer. Single-maintainer concentration risk is present; the project could stall if priorities shift.
• Agent Skills Specification dependency. Superpowers is built on the Agent Skills Specification standard. Changes to that spec (still evolving as of early 2026) could require framework updates. The spec’s allowed-tools field and progressive disclosure mechanisms vary in implementation across agents.