A diagram showing 14 skill cards being injected into a Claude Code session at startup
blog

Superpowers: Stop Claude From Vibe Coding

Skills, workflows, and guardrails for turning Claude into a disciplined coding agent. 🤖

Claude
Agent Workflows
Agent Skills
A diagram showing 14 skill cards being injected into a Claude Code session at startup

Superpowers is an open-source plugin that injects structured development skills — TDD, debugging, planning, code review — into Claude Code and other AI agents. Here's how it works.

Mini 4 min read

You open Claude Code. You type “build me a user authentication system.” Claude starts writing code immediately.

No questions about architecture. No clarifying what “authentication” means for your app. No asking if you want JWTs or sessions. Just code.

That’s the default. And for non-trivial work, the default is often wrong.

obra/superpowers is a plugin that changes this. It teaches Claude Code (and other AI agents) to stop, think, and follow disciplined software development workflows before writing a single line of code.


What Is obra/superpowers?

Superpowers is an open-source skills framework — a curated collection of 14 instruction sets that tell AI agents how to approach different development tasks. Think of it as a methodology plugin: you install it once, and it rewires how your agent handles everything from brainstorming new features to debugging production failures.

The core insight: AI agents are great at generating code, but terrible at knowing when to stop and ask. Superpowers defines explicit workflows for common development scenarios so the agent follows them automatically.


What Is a “Plugin” in Claude Code?

In Claude Code, a plugin is a packaged set of skills, hooks, and configuration that extends the agent’s default behavior. It gets injected into the agent’s context at the start of each session via SessionStart hooks — shell scripts that run on initialization and feed the skill instructions directly into the model’s working context. No manual prompting needed.

Each skill is a directory containing a SKILL.md file: a structured markdown document with trigger conditions and step-by-step instructions the agent must follow.

superpowers/
└── skills/
    ├── brainstorming/
    │   └── SKILL.md        ← "When user proposes a new feature, ask first"
    ├── test-driven-development/
    │   └── SKILL.md        ← "Iron Law: no production code before a failing test"
    ├── systematic-debugging/
    │   └── SKILL.md        ← "Four mandatory phases before writing any fix"
    └── ... (14 skills total)

Installing and Removing It

Claude Code:

# Install
/plugin install superpowers@claude-plugins-official

# Remove
/plugin uninstall superpowers

OpenCode (opencode.json):

{
  "plugin": ["superpowers@git+https://github.com/obra/superpowers.git"]
}

Cursor:

/add-plugin superpowers

⚠️ Note: Commands may vary by version. Check the repo README for the latest.


How Skills Activate (and Stop)

Skills activate automatically based on context. You don’t invoke them — a meta-skill called using-superpowers runs before every response and applies whichever skills match the current task.

What you doWhat activates
Propose a new featurebrainstorming — agent asks questions, writes design spec
Start a bug fixsystematic-debugging — four investigation phases required
Begin implementingtest-driven-development — blocks code before a failing test
Finish a branchverification-before-completion — final checklist runs

You can also explicitly invoke any skill with /skill [name].

To deactivate: uninstall the plugin. All skills stop applying immediately.


The 14 Skills

SkillPurpose
brainstormingDesign-first validation before any implementation
writing-plansDecomposes work into 2–5 minute atomic tasks
test-driven-developmentEnforces RED → GREEN → REFACTOR, blocks code-before-tests
systematic-debuggingRoot cause → pattern → hypothesis → fix phases
executing-plansManages task execution with verification gates
subagent-driven-developmentDispatches isolated subagents per task
dispatching-parallel-agentsCoordinates parallel execution with merge strategy
requesting-code-reviewStructures when and how to ask for review
receiving-code-reviewHow subagents verify spec compliance and code quality
using-git-worktreesCreates isolated directories per agent task
finishing-a-development-branchGuided decisions for merging or archiving
verification-before-completionFinal checklist before declaring a task done
using-superpowersMeta-skill: enforces skill-checking before any response
writing-skillsTDD applied to writing new skills

The Workflow in Practice

With superpowers installed, the same “build me an auth system” prompt now looks like this:

You:    "Build a user authentication system"
           ↓ brainstorming activates
Claude: "Before I start — JWTs or sessions? What's the user store?
         OAuth needed? Let me propose two approaches..."
           ↓ writes design spec, waits for approval
You:    "Approach 2, JWT, PostgreSQL"
           ↓ writing-plans activates
Claude: "Implementation plan: 8 tasks, each 2–5 minutes.
         Task 1: schema migration..."
           ↓ subagent-driven-development
Agents: [each task runs with TDD + code review gates]
           ↓ verification-before-completion
Done.   All tests pass. PR ready.

Is It Worth Installing?

For quick scripts or exploratory work: probably not. The upfront design phases feel like friction.

For anything non-trivial — features, APIs, refactors — superpowers forces the practices that keep AI-generated code testable and maintainable. It’s also a solid reference for writing your own SKILL.md files, since the 14 built-in skills are well-documented examples of what effective agent instruction sets look like.

Repository: github.com/obra/superpowers — MIT, by Jesse Vincent.

If you want to write your own skills from scratch, the Agent Skills series covers the format in depth.

Link copied to clipboard