AI codes better than me. Now what?

Robinson says he built a Rust image compressor that compiles to WebAssembly and a SvelteKit app without writing code by hand. He also used agents to create a Space Invaders clone in Verilog for an Analogue Pocket, despite having little experience with the hardware. Projects he would not have attempted a year earlier now feel achievable.

Robinson says current models are better at coding than he is and let him take on more ambitious work. He describes moving cursor.com from a CMS to Markdown and MDX in days, after expecting the migration might take weeks. As code gets easier to produce, he argues, the broader work of building software matters more.

Robinson expects models to become capable across most software engineering tasks, while arguing this does not mean software jobs will disappear. He recommends learning coding agents and using them to handle repetitive work such as boilerplate, JSON changes, and debugging shell errors. With code cheaper to produce, he says engineers should develop product judgment and build, market, and grow things people value.

Robinson cautions that relying too heavily on models can create the illusion of understanding. He suggests asking them questions like a patient pair programmer and using them to create guided mini-courses. For his recent projects, those lessons helped him learn Verilog, hardware abstraction layers, and compilation.

Robinson acknowledges that handing off code can feel like losing part of the craft. He says the goal was never code for its own sake, but building something worthwhile and being proud of it. He is optimistic about using AI to improve that work in 2026.