
A model will happily agree to almost any reasonable instruction. Tell it to write the test first - it agrees. Tell it to understand the cause before fixing - it agrees too. And five minutes into a long session it writes the implementation and adds the test on top.
That is not disobedience. An instruction in a prompt competes with the current context, with the tools, with the nearest goal - and usually loses. Superpowers tries to lift that argument off the level of persuasion and onto the level of a procedure that cannot be skipped.
Two hundred and seventy-two thousand stars - more than any other project in this field. Fourteen skills wired into one sequence, from the conversation about an idea to a merged branch.
What happens after you say what you want
The project describes its working cycle briefly, and it is worth walking through in full.
The agent does not start writing code. It steps back and asks what you are actually trying to do. Having teased a specification out of the conversation, it shows it to you in chunks short enough to genuinely read.
Once you sign off, an implementation plan is written. The phrasing in the README is precise and very telling: the plan must be clear enough for an enthusiastic junior engineer with poor taste, no judgement, no project context and an aversion to testing.
Then, on your word, subagent-driven development starts: each agent works through its task, its work is inspected and reviewed, and the cycle moves on. The project notes separately that the agent often works autonomously for a couple of hours at a stretch without deviating from the agreed plan.
Seven steps and their entry conditions
| Step | When it activates | What it does |
|---|---|---|
| Brainstorming | before code is written | draws out requirements by questioning, explores alternatives, presents the design in sections |