Plan Mode Is Dead

This recounts building a desktop coding app called Nuanced around the idea that explicit “plan mode” would become the central UI for human-AI software development. Plan modes were meant to do two things: give agents precise instructions and help humans form a coherent mental model of the system. Nuanced provided persistent chat threads, surfaced ambiguities, generated structured specs, and offered a Spec Tour; it also split work into plan and build modes so intent could be captured before implementation. The experience that motivated the product was rapid, opaque code generation: models could produce large amounts of code faster than humans could inspect or understand why decisions were made.

The implementation failed and revealed why plan modes are the wrong abstraction. Models became good enough at understanding repos that many explicit instructions were unnecessary, and long AI-generated specs were painful to read. The product enforced a linear plan→approve→execute workflow that conflicted with how thinking actually happens - an iterative understand→act→inspect→clarify→adjust loop that modern agents already embody. Forcing a plan artifact, separating planning from building, and asking users to choose modes added cognitive load. The core problem - keeping humans meaningfully informed as hundreds of agents change a system - remains unresolved; instead of long specs, agents must surface the few places where human attention has the highest leverage.