Superpowers

Jesse Vincent and the Prime Radiant team publish «Superpowers» under an MIT license; GitHub shows more than 130,000 stars.

The order of operations is the point. Before code, the agent runs a brainstorming phase on the real requirement, presents the specification in pieces short enough for a human to actually read and approve, then writes an implementation plan. Skills are markdown SKILL.md files across five domains: testing (strict red/green TDD, async tests, anti-patterns), debugging, collaboration, development (git worktrees, subagent dispatch) and meta-skills for authoring new skills. The core injects under 2,000 tokens at startup and a shell script finds and reads the rest on demand. That frugality keeps the agent's context from being spent up front.

A «mémoire conversationnelle SQLite» copies Claude transcripts outside the .claude directory to escape their automatic deletion after a month, indexes them with vector search, and generates summaries via Claude Haiku. «Claude Code» installs the project through its native plugin system; «Cursor», Codex and OpenCode need manual installation.

Dorra Bartaguiz shows why the question keeps returning. Code generated from one prompt was announced as 100% covered: she measured 33% overall coverage and a mutation score of 13.82%. Her term for everything wrapped around the model is the harness, and an agent is a model plus a harness. Skills are how that harness gets worked, with «Superpowers» and AXA's Skraft both resting on craft practices.

Vincent's own unfinished threads: pressure testing skills against extreme scenarios, extracting skills from technical books, applying persuasion principles to prompts.