Superpowers — Obra's Full Skills Framework for Coding Agents
Superpowers is the most-adopted third-party skills framework for coding agents. A set of composable skills covering code review, TDD, debugging, and refactoring workflows, designed to chain together inside Claude Code, OpenCode, and Codex.

Superpowers
Superpowers is obra's open-source skills framework — a curated set of composable skills (TDD, code review, debugging, refactoring, brainstorming) designed to chain together inside any agentskills.io-compatible coding agent. It is the most-adopted third-party skills framework outside Anthropic's own examples and pairs directly with the existing Superpowers 6: Parallel Agents concept page on this site.
When to use
- You want a code-review / TDD workflow that activates automatically on relevant triggers
- You're running a coding agent on a non-trivial repo and want a known-good skills baseline
- You maintain a team-standard procedure and want it expressed as a skills chain, not a doc
When not to use
- The framework is opinionated (TDD-first, spec-first) — if your team's process differs, you'll fight it
- You're on a tiny project where the skills' context overhead outweighs their value
How it works
Superpowers ships as a repo of individual SKILL.md files, each scoped to a single workflow step. The agent loads only the relevant skill on trigger, and the skills are designed to hand off cleanly to one another (e.g., a "brainstorm" skill's output becomes the input to a "write-spec" skill). This is the most explicit demonstration of the standard's progressive-disclosure design in the wild.
Install
npx skills add obra/superpowers
Full SKILL.md
Full source at the canonical repo: github.com/obra/superpowers. Embedded copy omitted at launch; the repo is a multi-skill tree and the canonical source is the right place to read individual skills with their current state.
Verify it yourself
Quick static check — clone to scratch, grep for the standard red-flag patterns, and confirm the allowed-tools for each individual skill is bounded. Superpowers is a collection; the right verification is per-skill, not per-repo.
Two-model review — for each individual skill you intend to use, paste that skill's SKILL.md (not the whole repo) into two different models with the Vetting Guide prompt. Pay attention to skills that touch Bash or Edit — these are the highest-capability skills in the framework and the most likely targets for hidden instructions.
Review transparency
| Field | Value |
|---|---|
| Last reviewed | 2026-08-28 (launch baseline) |
| Reviewers | site baseline (static scan only) |
| Source repo | github.com/obra/superpowers |
| License | MIT |
| Re-vet trigger | per-skill change in the repo; >30 days since review |
Related
- Superpowers 6: Parallel Agents — the concept page on this site
- Skills Collection Overview
- Vetting Guide
Related Articles & Guides
#skillsSkills Collection — Curated Agent Skills for Every Tool
A curated, vetted catalog of agent skills (SKILL.md) that work across Claude Code, OpenCode, Gemini CLI, Codex, Cursor, and 40+ other agents. Each skill includes install tabs, full source, and instructions for verifying it yourself.
#skillsHow to Vet an Agent Skill With Models
A repeatable methodology for vetting agent skills before installing. Static red-flag scan plus a copyable two-model review prompt that flags prompt injection, exfiltration, and over-broad permissions. Built on the Aug 2026 security research wave.
#skillsTaste-Skill — Anti-Slop Frontend Skill for AI-Generated Sites
Taste-skill (taste-skill-v1 and gpt-tasteskill) is the most-cited anti-slop frontend skill in the ecosystem. Pairs an opinionated design vocabulary with a set of hard 'no' patterns to keep AI-generated sites from looking generic.