I've noticed that the pushback usually falls into two camps: people who claim AI-generated code is complete garbage that takes longer to fix than to write from scratch, and those who are genuinely worried that we're forgetting how to actually think through a problem.
For me, the productivity gain is too massive to ignore. I’ve integrated tools like Claude Code and Cursor into my daily routine, and the speed at which I can move from a conceptual architectural sketch to a working prototype is insane. But I still catch myself hesitating to admit exactly how much of a specific function was generated versus hand-written during peer reviews. The "it's unusable" argument usually stems from a lack of proper prompt engineering or using the wrong tool for the job. If you're just pasting a vague request into a chat box and expecting a production-ready enterprise module, you'll be disappointed. But when you treat it as a collaborative pair-programmer, the experience changes.
To keep my skills sharp while maximizing speed, I've adopted a specific deep dive approach to AI-generated code:
-
The Logic Audit: I never merge a block of AI code without being able to explain exactly what every line does. If I can't explain it, I ask the LLM to break down the logic until I can.
-
Constraint-Based Prompting: Instead of asking "write this function," I provide the edge cases and constraints first.
-
Iterative Refinement: I use the AI to generate the skeleton, but I manually handle the critical business logic and security checkpoints.
Whether this is a "C vs. Assembly" moment is a great question. We didn't lose the ability to program when we stopped writing assembly for everything; we just moved up the abstraction layer. Using an LLM agent is essentially just another layer of abstraction.
The real risk isn't the tool itself, but the temptation to stop auditing the output. The "shame" vanishes the moment you realize that the value of a senior developer isn't in their ability to remember the exact syntax of a regex string, but in their ability to architect a system that actually works and scales. If an AI can handle the syntax, I can spend more time on the architecture.
Next Kiro Crew lets you ship full apps in five minutes flat →
these real-world AI monetization case studies, with plenty of directly applicable cases.