Superpowers

Executive Summary

Superpowers is a methodology-as-code for AI coding agents by Jesse Vincent (@obra, founder of Keyboardio). Rather than letting agents jump straight into writing code, it enforces a structured workflow — brainstorm → plan → TDD implementation → two-stage review — through 14 composable Markdown skills that trigger automatically.^[1] As of June 2026 the repo has 224,691 stars and 19,975 forks on GitHub — roughly 4x its February 2026 count — with the latest release (v5.1.0) shipped May 4, 2026 and pushes as recent as June 11, 2026.^[1] It is now distributed through Anthropic's official Claude plugin marketplace alongside seven other harnesses.^[2]

---

What It Is

Superpowers enforces a structured workflow instead of letting agents "just start coding":

1. Brainstorm → Refine rough ideas through Socratic questioning until a spec emerges
2. Plan → Bite-sized tasks with exact file paths and verification steps, written for "an enthusiastic junior engineer with poor taste, no judgement, no project context, and an aversion to testing"^[1]
3. Implement → Strict red-green TDD with subagent delegation
4. Review → Two-stage code review (spec compliance, then quality)
5. Finish → Merge, PR, or discard with worktree cleanup

The key insight: skills trigger automatically. You don't run commands — the agent checks for relevant skills before any task and follows them as mandatory workflows. Jesse reports agents working autonomously "for a couple hours at a time without deviating from the plan."^[1]

---

Platform Support

Coverage expanded sharply in 2026 — from four platforms in February to eight by June:^[1]

---

Skills Library (14 Skills)

Core Workflow

• brainstorming — Socratic design refinement before any code
• writing-plans — Detailed implementation plans with verification steps
• subagent-driven-development — Fresh subagent per task + two-stage review
• executing-plans — Batch execution with human checkpoints
• finishing-a-development-branch — Merge/PR decision workflow

Development Practices

• test-driven-development — Strict RED-GREEN-REFACTOR cycle
• systematic-debugging — 4-phase root cause process
• verification-before-completion — Prove it's actually fixed
• using-git-worktrees — Parallel development branches for isolation

Collaboration

• dispatching-parallel-agents — Concurrent subagent workflows
• requesting-code-review — Pre-review checklist
• receiving-code-review — Responding to feedback

Meta

• using-superpowers — Introduction to the skills system
• writing-skills — Create new skills (with TDD for skills)

---

Subagent-Driven Development (Key Innovation)

The most distinctive feature. For each task:

1. Dispatch implementer subagent (fresh context)
2. Implementer asks questions if needed
3. Implementer implements, tests, commits, self-reviews
4. Dispatch spec reviewer subagent → confirms code matches spec
5. Dispatch code quality reviewer subagent → approves or requests fixes
6. Mark task complete
7. Repeat for next task
8. Final code review of entire implementation

Why it works:

• Fresh subagent per task = no context pollution
• Two-stage review catches both spec drift and quality issues
• No human-in-loop between tasks = faster iteration
• Each subagent follows TDD naturally

As of v5.1.0 (May 2026), reviewer subagents dispatch as general-purpose agents with prompt templates — the dedicated code-reviewer named agent was removed, and legacy slash commands (/brainstorm, /write-plan, /execute-plan) were dropped in favor of invoking skills directly.^[1]

---

Releases Since February 2026

Development remains very active: the repo was pushed to on June 11, 2026, with 279 open issues.^[1]

---

Philosophy

From the README and Jesse's writing:

"Write tests first, always. Systematic over ad-hoc. Process over guessing. Simplicity as primary goal. Verify before declaring success."^[1]

The framework embeds persuasion principles (à la Cialdini's Influence) so agents follow workflows even under pressure scenarios — time pressure ("production is down!"), sunk cost ("but I already wrote it!"), and overconfidence ("I know how to do this"). Jesse uses subagents to pressure-test skills, simulating scenarios where an agent might try to skip steps.^[4]

---

Cautions

• Process overhead — the mandatory brainstorm/plan/review cycle adds ceremony that can feel heavy for trivial changes; reviewers note it shines on complex work, less on quick fixes^[5]
• Agents still cheat — enforcement is prompt-based, not mechanical; HN users report agents cutting corners on verification despite the skills^[6]
• Single-maintainer gravity — Jesse drives the overwhelming majority of commits; the bus factor is real despite the contributor count
• Churning API surface — v5.1.0 removed slash commands and the named reviewer agent, so downstream tools (e.g., Metaswarm) must track breaking changes^[1]

---

What Developers Say

"Jesse is one of the most creative users of coding agents (Claude Code in particular) that I know." — Simon Willison, creator of Datasette^[4]

"There is so much to learn about putting these tools to work in the most effective way possible. Jesse is way ahead of the curve, so it's absolutely worth spending some time exploring what he's shared so far." — Simon Willison^[4]

"The 'superpowers' set from Anthropic is really impressive." — oatmealsnap, Hacker News (recommending the brainstorming and git-worktrees skills; note the commenter misattributes authorship — it's Jesse Vincent's project)^[6]

On the critical side, HN user UncleEntity reports that Claude still tends to "cheat" on code-verification tasks even with skills installed, and that newer Claude Code versions found "really, really bad" issues when reviewing tests written under the workflow.^[6]

---

Related Work

• Metaswarm — Dave Sifry's multi-agent orchestrator builds directly on Superpowers for its foundational skills
• Anthropic Skills — the native SKILL.md format Superpowers popularized and extends
• BMAD Method — agile-persona alternative in the same agentic skills frameworks category
• Microsoft Amplifier — similar pattern of agents writing skills for themselves

---

Bottom Line

Superpowers is the most successful agentic skills framework in the wild — 224k stars in eight months, distribution on Anthropic's official plugin marketplace, and eight supported harnesses. It's not about giving agents more tools; it's about giving them discipline. The pattern it pioneered (automatic skill triggering, fresh-subagent isolation, two-stage review) is now the substrate other frameworks like Metaswarm build on.

---

Research by Ry Walker Research • methodology