When the same AI that writes your code writes your tests, the tests verify what the code does, not what it should do. Verification is the activity that changes most with AI, and the one we've adapted least.
Running Claude Opus and Sonnet in isolated git worktrees against the same Rust project reveals that LLMs have genuinely different analytical blind spots, and the union of both is better than either alone.
Every time AI generates code, it makes dozens of engineering decisions you didn't ask for. Dependencies, conventions, error handling. You didn't choose them, but you're stuck maintaining them.
AI doesn't save you planning time. It shifts where you spend it. And if you don't decide where, it decides for you.
One word can cut your responses from 800 to 90 tokens in any LLM. How to shift the model's probability distribution with a simple instruction.
AI coding assistants default to enterprise-grade complexity whether you need it or not. Designing before generating is what separates developers from prompt operators.
AI models solve 80% of benchmark problems but collapse to 23% when the task demands real comprehension. The gap between those two numbers is exactly what you bring to the table, and what you should never delegate.
There are two ways to get AI-assisted development wrong: letting it decide too much when you know nothing, and letting it ignore everything you do know. Both stem from the same problem.