Serious coding with LLMs. Lab notes 2026-10-27

Why it matters

This piece is the sixth lab note in an ongoing experiment testing the practical limits of LLM-generated code. The author is building a professional visual modelling application in C++, with nearly every line produced by Claude, alongside substantial documentation. The stated goal is neither to promote nor dismiss AI coding, but to probe where it breaks down using a serious, realistic use case. The central finding comes from applying the IfSQ code-inspection method, which revealed significant quality problems in the generated code, even though the LLM itself proved better than IfSQ at detecting certain kinds of issues. The author reports that fixing defects and maintaining quality now consumes more than half of his effort, reducing the productivity gains that make LLM coding feel impressive at first. He frames the broader problem with a taxonomy distinguishing incorrect output, correct but bloated output, and genuinely solid work, arguing that current tools tend toward the first two. His conclusion is that LLM agents cannot replace human engineers, though they remain useful under human direction.