Before I installed Superpowers, Claude would jump straight to code. I’d describe a bug, and instead of investigating the root cause, it would start rewriting functions. I’d ask for a new feature, and it would produce 400 lines of untested implementation without asking a single clarifying question. Sometimes the code was good and what I wanted. Often it wasn’t. The process was terrible either way.
Jesse Vincent fixed this with a plugin that doesn’t give Claude new capabilities. It gives Claude discipline.
What Superpowers Actually Is
Superpowers is a plugin for Claude Code that ships a set of “skills” - markdown files that encode development workflows. When Claude encounters a task, the bootstrap prompt - under 2,000 tokens - teaches it three things: you have skills, search for relevant ones before acting, and follow them when you find a match.
The skills aren’t suggestions. They’re structured decision trees that Claude reads and executes step by step. When I ask Claude to fix a bug, Superpowers intercepts the impulse to start coding and routes it through a four-phase debugging methodology instead. When I ask for a new feature, it runs a Socratic brainstorming session before anyone writes a line of code.
Jesse
Jesse Vincent built Request Tracker in 1996 - an open-source ticket tracking system that’s still in production at organizations around the world thirty years later. He created K-9 Mail , the Android email client that Mozilla eventually adopted and rebranded as Thunderbird for Android. He co-founded Keyboardio , which manufactures precision-milled ergonomic mechanical keyboards with open-source firmware. He co-founded VaccinateCA during the pandemic to help Californians find vaccine appointments.
Every project looks unrelated. The thread is toolmaking. Jesse builds things that give people structure for doing hard things - tracking bugs, managing email, typing comfortably, finding vaccines.
Now he runs Prime Radiant , an AI research lab building tools for what he calls “a world where agents do the work and humans do the thinking.” Superpowers is the flagship project.
He started writing about his coding agent methodology in a September 2025 blog post and formally introduced Superpowers a few days later . The timing was right - Anthropic had just rolled out Claude Code’s plugin system. Within months, Superpowers was in the official marketplace with over 143,000 installs, and Simon Willison was writing about it . Jesse kept shipping - version 4 in December split code review into two agents with formal review loops, and version 5 landed in March 2026.
The Skills That Changed How I Work
Superpowers ships thirteen skills across four categories - thinking, quality, collaboration, and infrastructure. Three changed how I work.
Brainstorming runs a Socratic design session before anyone touches an editor. Claude asks questions, explores alternatives, and saves the resulting design. I used it when I was trying to figure out how to run all of IntensityMagic’s business operations through Claude Code. The session asked questions I hadn’t thought through - should I build custom Google integration or use the MCP ecosystem? How should capabilities be distributed between users? What’s the right architecture for a two-person company that might scale to a hundred applications? I came in with a rough idea and walked out with a design spec.
I used the same skill when I needed a consolidated dashboard across all my apps. I typed /start INT-244 and said “brainstorm with me to figure out a good, modular approach.” Claude didn’t start writing React components. It asked what information belongs on a dashboard vs. in individual tools, how modular it needs to be when I add a new app, what’s the right level of density for something I’ll look at every day. The design session produced a spec. The implementation followed the spec. The dashboard works because the thinking happened first.
I used to think brainstorming with an AI was a novelty. Now it’s the first step on anything non-trivial.
Test-driven development enforces red-green-refactor without negotiation. If Claude writes implementation before a failing test exists, the skill forces a restart. I’d gotten lazy about this - TDD sounded right in theory, and I’d do it when I remembered. Now I don’t have to remember. The skill remembers for me.
Verification before completion requires running the actual commands and confirming output before claiming anything is fixed. I learned why this matters the hard way. During a long session, Claude’s context window compacted and it lost the step that said “run the tests.” It declared “Implementation complete!” after running type-check and lint - but never executed the unit tests. The code had bugs. I now have a structural gate that invokes this skill before any handoff. If the output isn’t in the response, the task isn’t done.
What I’ve built on top of Superpowers - the commit pipeline, the review triage, the project-specific slash commands - all depend on an agent that investigates before rewriting, tests before claiming success, and asks before acting. Superpowers is what makes that true every time, whether I’m paying attention or on to the next thing.
The skills are model-agnostic. My main agent runs on Opus for reasoning depth, review agents dispatch on Sonnet for efficiency, and quick lookups use Haiku. The methodology works the same way regardless of which model executes it.
The Marketplace
Superpowers spawned a small ecosystem. The marketplace now has nine plugins beyond the core - Episodic Memory for semantic search across past conversations, Elements of Style encoding William Strunk Jr.’s 1918 writing rules, Claude Session Driver for launching other Claude Code sessions as workers via tmux. The plugin system Jesse built turned out to be as generative as the plugin itself.
The GitHub repo has over 93,000 stars, with a major release roughly every two months since October 2025. Jesse’s still shipping. The skills keep getting sharper. And the layer between “very capable autocomplete” and “development partner with a methodology” keeps getting thinner.