Superpowers at 210K Stars: The Biggest Shift in AI Coding Isn't Smarter Models—It's Discipline

TL;DR

July 2026: A GitHub project called Superpowers gained 90K stars in 7 days, surpassing 210K total. It's not a new AI model, nor a code completion tool—it's a methodology framework that enforces software engineering discipline on AI coding agents.

Not a Tool, But a Methodology

If you've coded with Cursor, Claude Code, or Copilot, you've been here: AI starts writing code immediately, modifies three files before requirements are clear, and by the time you catch up, there are no tests, edge cases are ignored, and the commit message says "fix."

The model isn't the problem. It's missing engineering discipline.

Jesse Vincent, the author of Superpowers, saw this clearly. His open-source framework doesn't care how the model invokes tools or orchestrates tasks—it asks: Did your AI clarify requirements before writing a single line? Produce a design doc? Write tests? Make them pass? Pass review?

If AI won't do these by default, then force it.

The Seven-Stage Workflow

Superpowers breaks software development into seven sequential stages, each triggering a dedicated Skill:

Stage	Skill	Core Action
1. Brainstorming	brainstorming	Force vague ideas into clear specs; discuss alternatives; confirm design docs section by section
2. Git Worktrees	using-git-worktrees	Create isolated workspace on a separate branch; validate test baseline in a clean environment
3. Writing Plans	writing-plans	Break tasks into 2-5 minute chunks, each specifying file paths, expected outputs, and verification steps
4. Subagent Development	subagent-driven-development	One new subagent per task; independent execution, no context pollution
5. TDD Implementation	test-driven-development	Tests first—code that fails tests gets actively deleted by the agent
6. Code Review	requesting-code-review	Two-stage review: first check "doing the right thing," then check "doing it well"
7. Finishing Up	finishing-a-development-branch	Merge only after passing both review gates

These seven stages won't surprise any software engineer. Superpowers' value isn't in inventing new concepts—it's in turning these human-maintained engineering disciplines into the AI agent's default behavior.

Two Ruthless Mechanisms

Two-Stage Review

Every subagent output faces two gates. Gate one—direction review: does the implementation match the requirements? Scope creep? Gate two—quality review: is the code well-written? Security issues? Performance acceptable?

Fail either gate, and it's sent back for rework.

Evidence Gating

Superpowers enforces a hard rule: every assertion must have evidence.

"Should work." "Probably fine." "Seems okay."—these get rejected instantly. You must run tests and paste results. Screenshot proof. Print logs. Speak with data.

This is also why the project maintains a staggering 94% PR rejection rate. The maintainers hold themselves—and AI—to the same standard.

Getting Started

One command:

npx superpowers-zh --tool

Your AI coding tool (supporting 18 tools including Claude Code, Codex, and OpenClaw) instantly gains 20 Skills spanning from brainstorming to release. MIT licensed, completely free.

Search superpowers-zh on GitHub—the Chinese-enhanced fork has 4.4K stars of its own.

The Real Takeaway

Behind Superpowers' 210K stars lies a single truth: the barrier to AI-powered coding is no longer "knowing how to code"—it's "knowing how to manage AI with engineering methods."

In the first half of 2026, the AI tools ecosystem shifted its main theme from "model competition" to "infrastructure competition." The gap between models is narrowing, but the gap between workflow engines, review mechanisms, and test-driven pipelines is widening.

Superpowers' answer is simple: don't rush AI into writing code. Make it learn to think through a process first. This isn't about making AI smarter—it's about making AI follow the rules. And rules, after all, are the core of engineering.