Conceptual integrity and counting lines of code

The Spark: Why is an old debate suddenly relevant again?

"Lines of code is a terrible metric for productivity" — this is almost dogma in software engineering. Yet recently, veteran developer Simon Willison publicly came to its defense, arguing that in the age of AI coding agents, the metric actually deserves a second look. It sounds counterintuitive, but it cuts to the heart of how AI is reshaping software development: when the bottleneck is no longer writing speed, what is it?

The Core: A paradigm shift from 'writing fast' to 'thinking clearly'

Simon's argument has two key layers.

First, a change in magnitude changes the game. In the past, a top engineer writing 200 lines of production-ready code in a day was an exceptional achievement. Today, with AI agents, producing 1,000 lines of debugged, high-quality code is entirely feasible. This isn't a 10% improvement; it's a tenfold leap or more. When output undergoes such a radical shift, previously useless metrics can gain new meaning. It's like how you couldn't differentiate a good horse from another by "cargo capacity" in the horse-and-buggy era, but with trucks, "load capacity" becomes a key indicator.

Second, cognitive bandwidth is the new scarce resource. AI makes you 100 times faster at writing code, but your brain's ability to understand, track, and maintain that code hasn't scaled by the same factor. That's the real bottleneck. You can write a thousand lines a day, but you can't mentally hold the logic, dependencies, and potential pitfalls of a thousand new lines of code. Therefore, teams remain essential, and their core function shifts from "dividing the writing work" to "dividing the cognitive load." The value of a senior engineer is no longer in hand-writing every line, but in designing, reviewing, and maintaining the conceptual structure of the entire system, ensuring it doesn't become an inscrutable black box.

Trend Insight: Beware of 'Winchester Mystery House' Software

The most brilliant metaphor here comes from Claire Giordano: the Winchester Mystery House. The house has 140 rooms because its owner believed that as long as she kept building, the ghosts of those killed by her husband's rifles wouldn't haunt her. The result? A structure with no blueprint, chaotic layouts, dead ends, and bizarre rooms.

This perfectly captures the biggest risk of AI-driven development. When the cost of adding a new feature drops from "a week" to "an hour," the impulse to "just add one and see" becomes hard to resist. Software can bloat rapidly in the short term, sprouting strange features and modules that lack coordination, ultimately losing conceptual integrity — the unity, consistency, and predictability of its design. A system that has lost its conceptual integrity sees its future modification and maintenance costs grow exponentially, sometimes forcing a complete rewrite.

Practical Value: What is the new discipline for developers?

Previously, a significant part of development discipline was enforced by the "time cost" of implementation. A wild feature idea, estimated to take a week, would be naturally shelved as "not worth it." Now, AI compresses that "week" to an "hour," removing this natural filter.

Therefore, developers (especially seniors and tech leads) must establish a new, proactive discipline:

1. Design-first, not 'code first': Spend more time on architectural design, interface definitions, and module partitioning before using AI to generate code. Ensure new features align with the overall design intent.
2. Strict code review and refactoring: AI-generated code may be high quality, but does it match the style, patterns, and abstraction levels of the existing codebase? This requires more rigorous and aesthetically discerning reviews.
3. Embrace the courage to subtract: Resist the temptation of "it's fast, let's just add it." Regularly audit the codebase and decisively remove unnecessary, inconsistent "odd rooms."

Counterintuitive Takeaway: Redefining 'Seniority'

Perhaps the most counterintuitive insight from this piece is this: in the AI era, the work of junior engineers (writing specific code) may be heavily automated, but the core skills of senior engineers — system design, conceptual modeling, technical taste, and trade-off decision-making — are more valuable than ever. They are no longer the "fastest coders," but the "people who keep the entire system clear and comprehensible." For developers looking to enter or thrive in the AI era, investing in understanding the essence of software, architectural patterns, and systems thinking is more fundamental than simply learning how to use AI tools.