Superpowers

A skill library for coding agents by Jesse Vincent. It supplies the parts of a development process that a model will happily skip: explore the requirement before building, write the failing test first, plan before executing, debug from evidence rather than from a hunch, and verify before claiming something is done.

Why a library and not a longer prompt

A general instruction to "write tests" competes with everything else in the context window and loses, because it is one line against a task the model is actively trying to finish. A skill is loaded when the situation matches and occupies the foreground while it applies. The difference is not subtle in practice: the same model, on the same task, either writes the test first or does not.

The ones R21 leans on

• brainstorming — before any build. Its job is to surface the requirement the request did not state, which is where most rework comes from.
• test-driven-development — the failing test first, always. This is the one that catches "it works" claims that have never been executed.
• systematic-debugging — evidence before fixes. The alternative is changing something plausible and declaring victory when the symptom moves.
• verification-before-completion — the discipline of proving the thing rather than reporting it. R21's recurring failure mode is a green signal that means nothing, so this one earns its place more than the others.

The honest limitation

A skill library only works if the agent invokes it, and an agent under momentum is exactly the agent least likely to stop and load one. Every plausible rationalisation — "this is a simple question", "let me just look at the code first" — arrives precisely when the discipline would have paid.

That is a real gap and no library closes it by existing. R21 runs a standing convention alongside this one: before any non-trivial request, scan it against the skills already installed and say out loud which one fits, rather than quietly doing it by hand. The convention exists because the failure it addresses kept happening with the library already installed.

Verified against the GitHub API on 2026-08-22: MIT, 276,193 stars.