Software Is Recursive, and AI Is the Base Case

Summary

The essay presents software development as a recursive process increasingly carried out with AI assistance. A developer can ask Claude to prototype a tool for a task that has no existing solution, then deploy the imperfect prototype and use Claude again to address exceptions exposed by real-world data. In this view, each newly handled edge case moves part of the problem from an AI call into a more deterministic branch of the program, gradually expanding its decision tree. The author argues that software can be understood as a function over a problem space, with programming consisting largely of building a decision tree that returns the right result for each case. Claude makes early prototypes quickly because many ordinary cases are relatively easy, while the difficult remaining cases consume disproportionate effort. The essay describes this recursive expansion as a fractal-like process that continues until the domain is solved, the software is deprecated, or the project is acquired. It also compares Claude Opus, Sonnet, and Haiku: Opus can perform some tasks from a direct request, Sonnet may need a custom CLI, and Haiku may require a CLI, strong guardrails, and carefully designed incentives. Moving from Opus to Sonnet to Haiku to code can improve speed and reduce cost, but generally requires more human effort. The essay concludes that software may fundamentally involve managing this gradient between ideas, models, tools, and code.