// the find
obra/superpowers
An agentic skills framework & software development methodology that works.
Superpowers is a portable set of "skills" (markdown instructions + hooks) that impose a full SDLC on coding agents: brainstorm a spec, write a plan, work in an isolated worktree, implement via subagent-driven TDD, review, then merge. It ships install paths for over a dozen harnesses (Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Pi, and more), so it's for developers who already hand real feature work to an agent and want enforced process instead of ad-hoc prompting.
The harness portability is real engineering, not just marketing — the same skill definitions are wired into 15+ agents through per-harness install shims, so you're not locked into Claude Code. TDD is enforced structurally (RED-GREEN-REFACTOR, code written before a test gets deleted) rather than just suggested in a prompt. The diagnosing-superpowers skill reads back a session transcript with line-level evidence to explain why an agent went off the rails, which is a debugging loop most agent frameworks don't bother building. Skill behavior itself is tested via a separate eval harness (drill, from superpowers-evals) rather than taken on faith.
The brainstorming skill's visual companion loads a logo from primeradiant.com by default, which phones home the installed version every session unless you set SUPERPOWERS_DISABLE_TELEMETRY — opt-out, not opt-in, for something marketed as dev tooling. Despite being MIT and open, the README is explicit that they don't generally accept new skill contributions, so it's an install-and-use product more than a community-extensible framework. The full workflow (brainstorm → plan → worktree → subagent review → TDD → code review) is a lot of ceremony to sit through for a small fix, and there's no documented lightweight path for quick one-off tasks. Verifying that a skill actually triggers correctly means pulling in a second, separate eval repo rather than something self-contained in this one.