Member-only story
Claude already knows it should write a spec before touching code. It will tell you exactly that if you ask. But the moment you say “go build it,” it skips straight to implementation — every time. That gap between comprehension and compulsion is the exact problem Superpowers was designed to close, and it’s a problem that no amount of better prompting will fix.
The Problem Isn’t AI Skill — It’s AI Compliance #
Here’s the paradox that kicked off Superpowers: Claude can articulate test-driven development, spec-first design, and careful architectural planning with real sophistication. Ask it to describe the right engineering process and you’ll get a textbook answer. Tell it to actually build something, and all of that goes out the window.
Jesse Vincent, Superpowers’ creator, put it plainly: “When asked what he’d do, he gave the correct answer, but when told to actually do it, he skipped straight to implementation. Comprehension and compulsion are not the same thing.”
That sentence is the entire founding insight. The AI doesn’t lack knowledge of best practices — it lacks structural incentive to follow them. This is not a fringe problem. Claude Code now accounts for roughly 4% of all public GitHub commits as of April 2026 — an enormous volume…