Superpowers vs Halfcycle: a method you install, or one that keeps learning

dev.to2026-11-05 09:460 views

We make Halfcycle, so read this with that in mind. Superpowers is open source and well made; the full comparison, with every source, is on our site.

Superpowers, by Jesse Vincent and Prime Radiant, is a set of skills that teaches your coding agent a development method. If you have used it, you know it changes how the agent behaves: it stops jumping straight to code.

What Superpowers does well

Design before code. A written design for each piece of work, approved before building. Halfcycle does this too, and adds a product spec, the architecture and the phases above the per-component designs.
Test-first, with evidence. Halfcycle does this too: tests run before and after every change.
A fresh reviewer per task, and a final review of the whole branch. Halfcycle does this too, with each task reviewed twice by an agent that did not build it, and each finding recorded with its outcome.
The agent judges whether work is small or a project, and you can overrule it. Halfcycle sizes it from the parts of your system a change touches and tells you why: a small change, a mini phase or a full phase.
Open, readable, many agents. Sixteen coding agents named in its README. Halfcycle works with Claude Code and OpenAI Codex too, with OpenCode coming soon.

What Halfcycle adds

It learns from every project it runs on. When any project goes wrong in a new way, the cause and the fix reach yours automatically. What travels is the shape of the problem, never your code, and we do not train on your work. Superpowers' README describes no learning across projects.
Served, not installed. The method reaches your project one step at a time, so improvements arrive without you updating anything.
The running app gets tested. At the end of each phase an AI agent clicks through the live app in a browser. Superpowers' README describes no testing of a running app.
A console. Progress, phases, the system design, every agent and what it is building, and approvals recorded with who and when.
Risky actions are checked before they happen: deleting unsaved work, rewriting history, reading credentials, publishing.