Superpowers

The most popular AI coding tool of 2026 is not software. It is a methodology.

Open sourceAI coding agentsMITMethodology
GitHub stars200K+
LicenseMIT
Team size1

Superpowers is a set of fourteen Markdown skills that impose planning, test-driven development and code review on AI coding agents. Jesse Vincent released it on 9 October 2025; within months it had passed 200,000 stars and entered Anthropic's official plugin marketplace.

What's actually in it

There is no binary here. Superpowers is a collection of plain text instruction files that an AI coding agent reads before it starts work. One skill refuses to let the agent write implementation code before a failing test exists. Another forces it to reproduce a bug before theorising about the cause. A third walks a diff file by file with explicit stop-and-think points.

Skills activate automatically based on the task rather than through typed slash commands, which matters because the failure mode of agent instructions is that the model drifts away from them as context grows. Encoding the workflow as an enforced sequence rather than a remembered suggestion is the entire trick.

Why it exploded

It is free, MIT-licensed, and was released the same day Anthropic shipped its plugin system for Claude Code — a piece of timing that turned a personal tool into an ecosystem entry point. Adoption came almost entirely by word of mouth: no launch campaign, no staged press push, just engineers who tried it and found that the output held up.

It was accepted into Anthropic's official plugin marketplace in January 2026 and became the most installed plugin in that ecosystem. Other coding agents, including several competitors, now support the same skill files.

Vincent is a long-time open-source maintainer, a former release manager for Perl 5, and the founder of an email client later acquired by Mozilla. He had spent decades on the organisational side of software quality, then turned that experience on the newest problem in the field: agents that produce plausible code with no tests and no architecture.

The insight is deflationary and correct. Bigger context windows did not fix output quality; discipline did. A constrained model that writes forty working lines beats an unconstrained one that writes four hundred lines that almost work — and that is a lesson you can ship as text.

Estimates, not audited figures · Last updated: 2026-10-15

Related profiles

Keep exploring

110 profiles and counting — let fate pick the next one.

Frequently asked questions

What is Superpowers?

An open-source framework of Markdown skill files that enforce structured workflows — planning, test-driven development, review — on AI coding agents.

Does it cost anything?

No. It is MIT-licensed and free for personal and commercial use.

Who made it?

Jesse Vincent, a veteran open-source maintainer, released it in October 2025.