# Claude Superpowers Claude Superpowers is a [[Claude Code Plugins|Claude Code plugin]] by Jesse Vincent (Prime Radiant) that turns a coding agent into a disciplined software development methodology. It is not a feature pack; it is a workflow with opinions, enforced through composable skills that auto-trigger at the right moments. ## Core idea Instead of letting the agent "jump into code", Superpowers forces a brainstorm → spec → plan → execute → review loop. Skills are markdown documents that teach the agent *how* to do each phase, and they activate based on context — the user does not have to invoke them manually. The bet: agents are unreliable not because they lack capability but because they lack process. Wrap them in a methodology with strong defaults ([[Test-Driven Development (TDD)|TDD]], YAGNI, DRY) and they behave like a senior engineer instead of an enthusiastic intern. ## The workflow 1. **Brainstorming** — Socratic refinement of the rough idea before any code. Spec is shown back in chunks short enough to actually read. 2. **Git worktrees** — isolated branch + clean test baseline before work starts. 3. **Writing plans** — bite-sized tasks (2-5 min each), exact file paths, verification steps. Written for "an enthusiastic junior engineer with no judgement". 4. **Subagent-driven development** — fresh subagent per task, two-stage review (spec compliance, then code quality). Claude can run autonomously for hours without drift. 5. **Test-driven development** — strict RED-GREEN-REFACTOR; code written before its test gets deleted. 6. **Code review** — between tasks, by severity. Critical issues block progress. 7. **Finishing a branch** — verify tests, then merge / PR / discard. ## Why it matters - **Process beats prompting.** Skills encode *when* and *how* to do things. The agent stops needing the user to babysit each step. - **Subagents as units of work.** Each task gets a fresh context window with just enough information. Solves the long-context drift problem without giving up autonomy. See [[AI Subagents]]. - **Skills are the unit of teaching.** "Read the book and pull out reusable skills" is the bootstrapping pattern — turn any documentation into permanent agent knowledge. See [[AI Agent Skills]]. - **Cross-agent reach.** Distributed via the official Anthropic plugin marketplace and ports to Codex, Cursor, OpenCode, GitHub Copilot CLI, and Gemini CLI — same methodology across the agent ecosystem. ## Connection to my system Superpowers is the canonical reference implementation of the same pattern OSK uses: composable skills + agent identities + receptionist routing. Worth studying for what to steal (subagent-driven dev, plan-as-spec discipline) and what to consciously diverge from (Superpowers is dev-only; OSK is whole-life PKM). Related: [[Agentic Engineering]], [[Agentic TDD]], [[Software Design Patterns for AI Skills and Agents]], [[Claude Code Hooks]], [[AGENTS.md (File Convention)]]. ## References - GitHub: https://github.com/obra/superpowers - Release announcement (Jesse Vincent, 2025-10-09): https://blog.fsck.com/2025/10/09/superpowers - Anthropic marketplace: https://claude.com/plugins/superpowers ## Related - [[Claude Code]] - [[Claude Code Plugins]] - [[Claude Code Skills]] - [[AI Subagents]] - [[AI Agent Skills]] - [[Agentic Engineering]] - [[Agentic TDD]] - [[Test-Driven Development (TDD)]] - [[Brainstorming]] - [[Software Design Patterns for AI Skills and Agents]]