How I actually use AI when coding

I've noticed there's a gap between "AI can write code" and "how to actually use AI while coding." Most advice is either too abstract or too specific. Here are the patterns that actually work for me.

The IDE Assistant Pattern

The most useful thing? A good LLM in your editor. Not replacing you, but sitting beside you. Quick questions:

• "Write a React hook that handles form state"
• "Debug why this query is slow"
• "Generate a migration for this schema"

Fast feedback beats scrolling Stack Overflow every time. The key is speed — if it's faster than searching, it's valuable.

The Rubber Duck, But It Talks Back

Sometimes I just dump a confusing problem into Claude and ask "What am I doing wrong here?" Often the act of articulating it to AI is like debugging with a rubber duck, except the rubber duck occasionally has insights.

The Learning Tool

When I hit something unfamiliar — a new library, an API pattern, a database optimization — I'll ask AI to explain it. Not as gospel truth, but as a starting point for research. It accelerates learning because I'm already pointing in the right direction.

Where It Actually Fails

• Understanding your codebase's context: If it doesn't know your architecture, it'll suggest generic solutions.
• Making architectural decisions: Still all you.
• Security and performance: Never trust AI here without verification.
• New patterns you haven't seen: It can suggest, but you validate.

The Real Value

The pattern I've found is this: AI is great for the known unknowns. You know what you need to build, you just need to translate it quickly. That's where it shines.

Don't use it to figure out what to build. Use it to help you build faster.