Superpowers by Jesse Vincent is a plugin for Claude Code (also works with Cursor, Codex, Copilot CLI, Gemini CLI). It takes the usual “throw a prompt at the agent and pray” loop and wraps it in an actual workflow: brainstorm, plan, write tests, implement, review, repeat. Each step is a skill the agent picks up on its own when the context fits.

The clever bit is auto-activation. No /brainstorm, no /plan. The agent checks the skill registry before starting and pulls in whatever matches. TDD kicks in when it’s time to write code. Code review runs between tasks. The skills are labeled as mandatory, not optional, which matters — otherwise the model will cheerfully skip straight to typing.

Four principles in the README: TDD, systematic over ad-hoc, complexity reduction, evidence over claims. That last one is the one I keep coming back to. It forces the agent to actually verify things before calling them done, and “declared done but wasn’t” is the failure mode that eats the most of my time.

Subagent-driven development with git worktrees is what lets the thing run unattended. Tasks dispatched to isolated branches, two-stage review against the spec, orchestrator only wakes you up when something genuinely needs a human. The author says a couple of hours of autonomous work is normal. I believe it, structurally — but I’d still want to read the diff before trusting any of it with production code.

Install on Claude Code with /plugin install superpowers@claude-plugins-official. MIT licensed.