How AI Coding Tools Actually Changed My Workflow
Not a hype piece. After a year of daily use, here's what AI coding assistants are genuinely good at, where they waste your time, and the one workflow shift that made the biggest difference.
I've been using AI coding assistants daily for over a year now. Long enough to have moved past the initial excitement and also past the inevitable backlash phase. Here's where I actually landed.
What they're genuinely good at
Boilerplate and scaffolding. Setting up a new API route, writing a TypeScript interface from a JSON shape, generating a form with validation — tasks that are mechanical and time-consuming but don't require deep problem-solving. This is where AI assistance pays back immediately.
Searching unfamiliar APIs. Instead of tabbing to the docs every five minutes, I can ask "what's the syntax for a Firestore composite query with an inequality filter" and get a working example in context. I still verify against the docs, but the first draft is usually close enough to be useful.
Explaining code I didn't write. Open a file I haven't touched in six months and ask what a particular function is doing. Faster than re-reading it cold.
Where it wastes time
Overconfident wrong answers. The model will generate code that looks completely plausible but contains a subtle bug or uses a deprecated API. If you're not experienced enough to catch it, you can end up spending more time debugging AI output than you would have spent writing it yourself.
Large refactors. The further a task gets from a self-contained chunk, the worse the output gets. "Refactor this entire module to use a different data model" usually produces something that compiles but breaks the surrounding context in ways that aren't obvious.
Architecture decisions. AI will give you an answer to any architecture question you ask. That answer will sound confident. It's frequently not the right answer for your specific constraints. Use it for ideas, not decisions.
The workflow shift that actually helped
Stop using it as an autocomplete that runs in the background. Start using it as a pair programmer you consult deliberately.
I write the structure and the logic myself. When I hit something mechanical — writing out 20 similar test cases, generating TypeScript types from a schema, stubbing out a component — I switch to the AI. When I'm done with the mechanical part, I read everything it wrote before accepting it.
That combination — me driving, AI handling the tedious parts — is where I get the real productivity gain. Letting it drive while I watch is where I lose time.