superpowers: 179K Stars for an Agentic Skills Framework That Makes AI Coders Actually Work

If you’ve used AI coding assistants long enough, you’ve hit this annoying problem: every new session, it’s like onboarding an amnesiac employee. You re-explain project conventions, code style, workflow preferences — again and again. obra’s superpowers, with 179K stars, targets exactly this pain point.

What This Project Is

superpowers is an agentic skills framework open-sourced by obra (Jesse Vincent, a veteran developer and active figure in the Claude Code community). In one sentence: it’s a “skill pack + methodology” that gives AI coding assistants (like Claude Code) a reusable, composable skill system.

Think of it as writing an “employee training manual” for your AI assistant. Things you used to explain verbally every time now become structured skill files that the AI loads automatically when relevant tasks come up.

Core Ideas

Skills as files. Each skill is a Markdown file describing a specific capability: how to do TDD, how to write commit messages, how to debug production issues, how to review code. The AI invokes them at the right moments.

Built-in methodology. The project isn’t just a tool — it ships with a full software development methodology. For example, it requires the AI to write tests before implementation, mandates self-review after completing tasks, and demands it ask questions instead of guessing when uncertain. These “rules” used to require constant prompt repetition.

Composition over accumulation. Skills can reference and compose with each other. An “implement new feature” skill can internally reference three sub-skills: “TDD process,” “code conventions,” and “commit standards.”

My Real Experience

The biggest change is stability. Before, asking AI to write features felt like a lottery — sometimes brilliant, sometimes a mess. After installing superpowers, behavior patterns stabilized: it runs tests first, runs them again after writing, and walks through a self-check before committing.

New sessions no longer start from zero. Project conventions, tech stack agreements, and team habits live in skill files. New sessions load them automatically, eliminating repetitive explanations.

There’s a learning curve. You need to think through your team’s workflow before writing good skills. This framework is for people willing to “codify first, reuse later.”

Quick Start

Per the README, the core is placing the skills directory in your project or user directory:

git clone https://github.com/obra/superpowers.git
cd superpowers
# Follow the README to link skills into Claude Code's skills directory

Then use Claude Code as normal — it automatically recognizes and loads skills. Start with the bundled skills, learn the structure, then write your own.

Pros and Cons

Pros:

• Solves the core pain points of AI “amnesia” and unstable behavior
• Skills are just Markdown — no barrier to writing them
• Mature methodology with built-in best practices (TDD, self-checks)
• 179K stars, active community, rich examples
• Deep integration with the Claude Code ecosystem

Cons:

• Primarily built around Claude Code; limited adaptation for other tools
• Building your own skill system requires upfront investment
• Shell-script assembly can hit minor environment issues on Windows
• With many skills, the AI occasionally picks the wrong one to load
• Not much help for pure chat scenarios — it’s oriented toward coding tasks

Comparison

Approach	Form	Reusability	Learning Curve	Best For
superpowers	Skill file framework	High	Medium	Long-term projects, team conventions
Handwritten CLAUDE.md	Single-file conventions	Medium	Low	Simple per-project rules
Fresh prompts each time	Verbal instructions	Low	Low	One-off tasks
Cursor Rules	Rule files	Medium	Low	Cursor users
MCP servers	Tool services	High	High	External tool integration

superpowers and CLAUDE.md don’t conflict — the latter defines “what the project is,” the former defines “how work gets done.”

Who Should Use It

1. Heavy Claude Code users — want stable, controllable AI behavior
2. Team leads — want to codify team conventions for AI to follow
3. Methodology enthusiasts — believe “codified process beats improvisation”

After three weeks, my clearest observation: the rework rate on AI-written code dropped noticeably. Before, I often had to say “no, redo it.” Now it’s more often “yes, keep going.” For that alone, it’s worth the time.

---

About the Author

Liudingyu is a full-stack developer and heavy GitHub user. With 900+ starred repos over the past 3 years, this site only covers tools I’ve actually used or deeply researched.

📧 Found a great tool to recommend? Email [email protected]