I Asked Claude to Interview Me About AI in Software Engineering

Where does AI coding assistance actually pay off, and where does it quietly cost you more time than it saves? I let Claude generate four hard questions about using LLMs in software Where does AI coding assistance actually pay off, and where does it quietly cost you more time than it saves? I let Claude generate four hard questions about using LLMs in software engineering, then answered them on camera. No script, no filter.

I cover greenfield scaffolding vs. legacy code archaeology, why test coverage requirements should go up with AI-generated code (not down), the senior-vs-junior “feels faster” trap, and where the real boundary sits between AI autonomy and human architectural decisions.

Topics covered:

Where LLMs shine: greenfield scaffolding, infrastructure automation, boilerplate elimination, understanding unfamiliar/legacy codebases
Where LLMs quietly cost more time: unclear problems, overly complex multi-step agentic workflows
Why AI-generated code should raise your testing bar, not lower it: behavior-based tests, mutation testing, security and performance test suites
The perception-vs-reality trap
Who should own architectural decisions when an AI can generate the alternatives

If you’ve felt like you’re moving fast with AI but aren’t sure you actually are, this one’s for you.