superpowers: Claude Code's 118K-star autopilot plugin

A GitHub repository called superpowers, created by developer Jesse Vincent (obra), just crossed 118,000 stars — making it one of the fastest-growing AI developer tools on the platform. On March 27, 2026 alone, it gained 2,993 new stars in a single day. The reason for the explosion of interest: superpowers solves a problem every AI coding tool user has run into — the AI starts strong, then gradually drifts off course, breaks its own plan, or needs constant hand-holding.

Superpowers gives Claude Code (Anthropic's coding agent), Cursor, Codex, OpenCode, and Gemini CLI a structured methodology — a discipline framework that forces the AI to plan first, write tests before writing code, and review its own work before declaring success. According to the creator: "It's not uncommon for Claude to be able to work autonomously for a couple hours at a time without deviating from the plan you put together."

The Problem It Solves

If you have used Claude Code or Cursor to build anything beyond a simple script, you have probably hit this pattern: you give the AI a task, it makes good progress, then somewhere around the 30-minute mark it starts going in circles, makes a change that breaks something it already fixed, or loses track of the original goal and starts improvising. You end up supervising the AI more than you expected — constantly correcting, redirecting, and re-explaining.

This happens because AI coding agents, without external structure, tend to be reactive: they respond to the immediate prompt but lack a persistent methodology for how software should actually be built. Superpowers provides that methodology.

The framework enforces a sequence that professional software engineers follow: brainstorm → plan → write tests first → write code to pass the tests → review. By making the AI commit to this sequence before touching any code, superpowers creates a feedback loop that self-corrects — the AI cannot declare success until the tests actually pass and the review is complete.

How It Works — The Four-Phase Workflow

Superpowers structures every coding task through a four-phase process:

• Phase 1 — Brainstorm: The AI explores the problem space, asks clarifying questions, and identifies edge cases before writing a single line of code. This prevents the common failure mode of AI starting to code before fully understanding the requirement.
• Phase 2 — Plan: A detailed written plan is produced and agreed upon before implementation begins. The plan acts as an anchor — when the AI drifts, it can be brought back to the plan rather than starting over.
• Phase 3 — Test-Driven Development (TDD): Tests are written before the implementation code. The cycle follows the RED-GREEN-REFACTOR pattern: write a failing test (RED), write minimal code to make it pass (GREEN), then clean up (REFACTOR). This means the AI cannot fake its way to completion — either the tests pass or they do not.
• Phase 4 — Review: Before the task is considered done, the AI reviews its own work against the original plan. This catches the subtle drifts and second-order bugs that appear when AI optimizes for the immediate task but forgets the larger context.

Superpowers also uses git worktrees (isolated copies of a repository that let the AI work on a task without touching the main codebase until the work is verified) and subagent-driven development (spawning specialized parallel agents for different parts of a task). Both techniques significantly reduce the chance that a failed experiment corrupts your working code.

Who It Works With

Superpowers is not Claude Code-specific — it works across five major AI coding platforms:

• Claude Code (recommended, easiest install via the official plugin marketplace)
• Cursor (via CURSOR.md configuration)
• Codex (OpenAI's coding agent)
• OpenCode (open-source coding agent)
• Gemini CLI (Google's terminal-based AI tool)

The core methodology is the same across all platforms — only the installation step differs.

How to Install in Claude Code

The fastest installation path is through Claude Code's official plugin marketplace:

/plugin install superpowers@claude-plugins-official

After installing, you can activate the framework by starting any task with the brainstorm phase:

/brainstorm Build me a REST API for user authentication with JWT tokens

The AI will walk through all four phases automatically. You review and approve the plan, then it builds and tests without you needing to intervene until completion.

Why 118,000 Developers Starred This in Under 90 Days

Superpowers gained most of its stars in the first 90 days of 2026, coinciding with a shift in how developers are using AI coding tools. The early adopter phase — where AI helps you write a function here, fix a bug there — has given way to people trying to use AI for entire feature branches and multi-hour development sessions.

Those longer sessions exposed the drift problem at scale. Superpowers landed at exactly the right moment with a principled solution, and the 118,000 stars reflect how widely that problem is felt.

If you are spending more time correcting AI than building with it, this is the framework to try. The full repository, installation instructions for all platforms, and detailed methodology documentation are at github.com/obra/superpowers.

Related Content — Get Started with Easy Claude Code | Free Learning Guides | More AI News