Superpowers
The kit of skills, sub-agents and commands by Jesse Vincent that turns Claude Code from a generalist agent into a disciplined team. Full credit to the author — I'm just a user.
What it is
Superpowers is a kit of files you drop into `.claude/` that adds three things to Claude Code: skills (reusable expertise memos), sub-agents (specialised roles with their own context and own tool perimeter), and slash-commands (orchestrated workflows).
The side effect, perhaps the most important one, is that Claude Code stops answering everything by itself. It calls the right skill, dispatches to the right sub-agent, follows the workflow. Discipline no longer lives in the developer's memory — it lives in the repo.
How it works
A Superpowers sub-agent is a markdown file with a frontmatter that describes its name, model, tools and mission. Example, my CSO (Chief Security Officer) agent that I invoke on anything sensitive:
---
name: cso
model: opus
tools: [Read, Grep, Glob]
description: Security review — IDOR, RLS, CSP
---
Mission: challenge the security risks of the PR.
When I trigger /call-tech-lead, the orchestrator dispatches several sub-agents in parallel (security, tests, perf, accessibility) that respond WITHOUT seeing each other's opinions in the first round. Only then does the tech lead compare opinions, detect disagreements, arbitrate. It's convergence after debate, not convergence by mimicry.
Depth routing is explicit: lean for an understood bug or a small refactor, standard for a normal feature (the default), full for sensitive topics — auth, payment, PII, AI, pricing. If the task is trivial, no sub-agent: direct edit.
How to install it
Clone the kit from GitHub and drop it into the project (or globally in `~/.claude/`):
git clone https://github.com/obra/superpowers ~/superpowers
cp -R ~/superpowers/.claude/agents .claude/
cp -R ~/superpowers/.claude/skills .claude/
cp -R ~/superpowers/.claude/commands .claude/
On the next Claude Code launch, the skills and sub-agents appear automatically. Commands are invoked via /command-name.
How I use it daily
On this very qaconsulting.fr site, for example, every implementation plan goes through: skill `brainstorming` to scope, skill `writing-plans` to produce the plan, skill `subagent-driven-development` to execute, and a final review by a dedicated sub-agent before the PR.
The benefit isn't just better quality — it's also that I no longer need to remember the steps. The repo remembers for me. And another developer who picks up the project has access to the same tooling I do, with no onboarding.
Sources & credits
Superpowers is designed and maintained by Jesse Vincent (GitHub handle: `obra`). All credit for the design of skills, sub-agents and commands goes to him. This article reflects my personal usage.