My Honest Take on Building With LLMs

I've been experimenting with LLMs for developer tooling. Not chatbots or content generation - actual workflow improvements. Here's what I've found.

What Works

Code review assistance works well when you constrain the scope. Ask an LLM to check for naming consistency or find potential null pointer issues, and you get useful feedback. Ask it to "review this code" and you get generic advice.

What Doesn't Work (Yet)

Architecture suggestions. LLMs don't understand your project's context deeply enough to make good architectural decisions. They can suggest patterns, but they can't weigh the trade-offs specific to your situation.

What I'm Still Figuring Out

The line between "useful automation" and "unnecessary complexity." Adding an LLM to a workflow adds a dependency, latency, and a new failure mode. Sometimes a simple script is better.