A frontend engineer built a working programming language called Cutlet over four weeks using Claude Code, without reading any of the generated code. The post covers the language's features (vectorized operators, prototypal inheritance, mark-and-sweep GC), the methodology behind the experiment, and four key skills for agentic engineering: problem selection, communicating intent via spec documents, creating a rich agent environment (test suites, sanitizers, linters, Docker with full permissions), and optimizing the agentic loop. The author reflects on authorship and credit, mental health risks of addictive AI tooling, and concludes that software engineering skills remain essential but will transform significantly.
Table of contents
A tour of CutletWhy build this?Four skills for agentic engineeringUnderstanding which problems can be solved effectively using LLMsCommunicating intentCreating an environment for the agent to do its best workOptimizing the agentic loopIs software engineering as we know it dead?Is it fair to take credit for Claude’s work?This wasn’t good for my mental healthWhat do we do with these new superpowers?What’s next for Cutlet?Sort: