A leash for your coding agent that actually makes it behave
Superpowers is a plugin framework that forces AI coding assistants to plan, test, and review before they ship broken code.

What it does
Superpowers installs as a plugin across eight AI coding tools — Claude Code, Cursor, Codex, Copilot CLI, and others — and overrides the default “immediately start typing code” behavior. It makes your agent interview you about intent, write a design doc in digestible chunks, get your sign-off, produce a granular implementation plan, then execute via subagents with mandatory test-driven development and two-stage code review.
The interesting bit
The framework treats your AI like “an enthusiastic junior engineer with poor taste, no judgement, no project context, and an aversion to testing” — and builds guardrails accordingly. Skills trigger automatically; the agent can’t skip steps. The README claims Claude can run autonomously for hours without drifting from plan, which is either impressive or terrifying depending on your last agentic coding experience.
Key highlights
- Mandatory workflow, not suggestions: Agent checks for relevant skills before any task; critical review issues block progress
- RED-GREEN-REFACTOR enforcement: Writes failing test first, watches it fail, writes minimal code, commits; deletes any code written before tests
- Subagent-driven development: Fresh subagent per task with spec-compliance review, then code-quality review
- Git worktree isolation: Each feature branch gets its own workspace with verified clean test baseline
- Cross-harness portability: Same skill set works across Claude, Cursor, OpenAI Codex, Gemini, GitHub Copilot CLI, and others
Caveats
- New skill contributions are generally not accepted, and any skill updates must work across all supported agents — high bar for community extension
- Installation and updates are “somewhat coding-agent dependent” and vary significantly between tools
- The 216K stars figure appears suspiciously high for a methodology repo; verify independently if that number matters to you
Verdict
Worth installing if you’re already using AI coding agents and tired of watching them confidently refactor your working code into a broken mess. Skip it if you prefer tight manual control or your workflow doesn’t involve any of the supported harnesses.