This repository follows the Agent Skills spec, so one command installs it into your AI client — works across Claude Code, Codex, Cursor, Copilot, Windsurf, and 18+ other clients.
npx skills@latest add obra/superpowers
This repository bundles 14 skills, and you can install just one:
9 more are not listed — run npx skills@latest add obra/superpowers --list to see them all.
Add -g install globally (available in every project) · add --agent claude-code to target a client · add --yes to skip the prompts. Restart your AI client afterward and it will pick it up.
OverviewOverview
Key Features
The foundational agent engineering governance and software craftsmanship framework (superpowers). Key capabilities include: 1. pre-execution guardrails (enforces requirement clarification and architectural brainstorming before modifying code to prevent hallucinated changes); 2. test-driven development (TDD) loop (mandates writing failing unit tests before feature code to guarantee logic coverage); 3. anti-degradation refactoring standards (enforces surgical micro-edits and crafts concise, semantic git commit messages).
No description beats seeing it actually run — the clips below are exactly that.
Superpowers: NEW Spec Toolkit Ends Vibe Coding! 100x Better Than Vibe Coding (Full Tutorial)
WorldofAI10:5997.1k views6mo ago
This video provides an in-depth walkthrough of Superpowers, a spec-driven development toolkit engineered for AI coding agents. Addressing the core pitfalls of unconstrained vibe coding—such as brittle architectures, hallucinated logic, and compounding technical debt—the tutorial demonstrates how to author formal specification documents, establish strict architectural guardrails, and implement test-driven constraints before agents generate code. The presentation details automated validation loops, continuous requirement verification, and structured plan execution across complex codebases. It illustrates how enforcing rigorous engineering discipline minimizes model drift, eliminates runaway hallucinations, and ensures verifiable long-term software maintainability.
The videos are published by third parties on public video platforms; they are not hosted here and are not maintained by the project authors. View counts and durations are platform data. Clicking a cover opens the platform in a new tab to play the video; watching requires access to that platform.
Ratings store no account and no plaintext IP — the IP participates only as an irreversible hash in the one-vote-per-IP check, and never appears on any page.