Lines of code are a poor metric for programmer productivity, argues a Hacker News post. The author counters claims that Large Language Models (LLMs) dramatically boost output by generating vast amounts of code, pointing to the long-standing view that LOC measures effort poorly and is poorly correlated with software quality, defects, or time. Programming, the post emphasizes, is fundamentally about managing complexity and representing ideas, not just writing code quickly. This perspective is reinforced by the preface to the first edition of SICP, which states programs must be written for people to read and that the core challenge is controlling intellectual complexity in large systems.

LLMs accelerate code generation but risk rushing implementation, the author contends. While LLMs can speed up the part of programming that involves writing code, they often push teams towards implementation too early. The tooling and agent harnesses encourage generating output (code, plans) rapidly, potentially locking in designs prematurely. This creates a false sense of progress and makes future changes harder. The high fidelity of generated code makes it feel final and important, hindering the disposable nature of early design sketches and whiteboard scribbles that allow for freer iteration.

Premature code focus harms maintenance and understanding, the post concludes. LOC's impact on maintenance is significant, though hard to quantify precisely. Generated artifacts, even text-based ones like ASCII art or PlantUML, feel more concrete and weighty than ephemeral plans, leading to confusion about their true value. The ease of iterating on low-fidelity prototypes is lost when code becomes the primary output. Ultimately, the author warns that LLMs may reinforce the misconception that LOC matters, distracting from the essential work of design and complexity management that defines good programming.