JCConf 2026

"With AI, Code doesn't matter anymore." It's a seductive idea, and it's gaining believers: if specs and tests are the real source of truth and code is just a disposable artifact the model regenerates on demand, then technical debt stops being a problem. If the code has become impossible to change, just let AI rewrite it all from scratch. Clean architecture, naming, and separation of concerns have become yesterday's rituals — like punch cards.
This talk argues the opposite: internal quality matters more in an AI-driven workflow, not less. There are technical reasons why LLMs are more sensitive to well-structured code than humans are, and we'll look at examples drawn from real experience, alongside research papers that shed some light on the matter.
It's also a question of economics. If technical debt can be measured in the tokens — that is dollars — needed to work around it, then keeping it low directly saves money.
But there's a catch: "code quality" for an LLM is not exactly the same thing as what computing literature has traditionally called code quality. Modularization and good naming matter even more than before, but some classic "clean code" advice can actually worsen a model's performance. We'll look at where the two notions diverge, and how to improve your productivity scientifically — by measuring the effect of good structure and good instructions rather than assuming it.
Main Takeways:
Why LLMs are technically more sensitive to code structure than human readers
Which conventional "clean code" practices can backfire on a model, and why
How to measure and improve assistant productivity scientifically, rather than by gut feel