superpowers

The Superpowers GitHub repository is an agentic skills framework that has quietly become one of the most-starred projects in GitHub history — more than 306,000 stars — by fixing an unglamorous problem: AI coding agents write code quickly, but not always code you can trust. Superpowers is not a model, a CLI or an IDE. It is a bundle of plain-Markdown skills plus a working methodology that installs into the agent you already use and imposes professional discipline on every task. The Superpowers GitHub project, created by Jesse Vincent and the Prime Radiant team, turns “the agent jumped straight into editing files” into “the agent asked, planned, tested, then built”. The official home of the project is the obra/superpowers repository on GitHub, where you can browse the source code, releases and issue tracker.

What it is

Superpowers has two halves. The first is a library of composable skills — structured instruction files an agent loads for specific jobs — covering the full software development lifecycle: brainstorming, writing plans, test-driven development, systematic debugging, subagent-driven development, requesting and receiving code review, using git worktrees, and even a skill for writing new skills. The second half is the methodology that binds them: instructions that make the skills mandatory workflows rather than optional suggestions, triggered automatically from the task description. Because the skills are plain text, they are portable across Claude Code, Codex, Cursor, Gemini CLI, OpenCode and Copilot CLI — the same discipline follows you from tool to tool, and the whole framework is MIT-licensed.

Key features

• Automatic triggering — skills fire from the task itself; there are no commands to memorise and no slash-command ritual.
• Brainstorming before building — the agent draws out intent, plays a spec back in readable chunks and gets your sign-off before writing code.
• Test-driven development as the default — implementation happens under strict red-green-refactor instead of test-after-the-fact.
• Systematic debugging — a structured root-cause process replaces random guess-and-check editing.
• Subagent-driven development — larger work is split across subagents with review steps between them.
• Composable and hackable — skills are plain Markdown files; teams edit, drop or add skills to match their own standards.

Who it is for

Engineering teams adopting AI coding agents who have discovered that raw speed creates review debt. Team leads get a process that runs identically at 9am and 11pm, for seniors and juniors alike; solo developers get the habits of a disciplined team without hiring one. Agencies shipping client work with agents — our situation — are the sweet spot: the methodology is effectively your internal process, written down where the agent cannot ignore it. It is less relevant if your agent use is limited to autocomplete and one-line edits; the framework earns its keep on multi-file, multi-step work where an undisciplined agent can wander expensively.

Getting started

Install Superpowers into your agent using the per-agent instructions in the Superpowers GitHub README — for most agents it is a plugin or marketplace install that takes minutes, and the repository itself doubles as the documentation. Then run one real task end to end: describe a small feature and watch the sequence — brainstorming questions, a spec for your approval, a written plan, then test-driven implementation with review checkpoints. Resist skipping the spec step on that first run; the friction is the product. Afterwards, read the skills that actually fired and edit one to match your team’s conventions — that customisation loop is where the long-term value compounds.

Our take

Superpowers is the most transferable idea in agent tooling this year, because it treats process — not prompts — as the asset. The honest caveats: the methodology adds ceremony, and on trivial tasks the spec-plan-test sequence can feel heavy; the skills also encode one team’s opinions, so expect to fork and adjust them rather than adopt them blindly. But the direction is unquestionably right, and the framework’s portability means the investment survives your next tool switch. We analysed the workflow in depth on the MTD blog — read our full deep-dive — including where each skill earns its place. If your agents write production code, the Superpowers GitHub repository is required reading.