Building with AI: What Actually Changes When the Tools Get Good

A year ago I started treating AI tools as a genuine part of my development workflow rather than a novelty to experiment with. Not as a search engine replacement, not as an autocomplete upgrade — as a collaborator that handles a real share of the thinking.

Here's what actually changed.

The Bottleneck Moved

Before: the bottleneck was typing. I knew what I wanted to build, I just had to write it. AI tools mostly eliminated that bottleneck for code that fits recognizable patterns — CRUD operations, API integrations, component scaffolding, test boilerplate.

After: the bottleneck is clarity. The tools are fast enough that the constraint is now how well you can specify what you want. Vague prompts produce vague code. Precise prompts produce precise code. The skill that matters most is articulating intent.

This sounds obvious in retrospect. But it means the job description changed. Less typing, more thinking-before-typing.

What Didn't Change

Judgment. The tools don't know what's worth building, what tradeoffs are acceptable, what the user actually needs versus what they asked for. Every output needs a reviewer who understands the context — and that reviewer is still you.

Debugging complex state. When something is subtly wrong across multiple systems, AI assistance helps but doesn't replace the mental model you need to hold in your head.

Architecture decisions. "How should this be structured?" is a question where AI can surface options and tradeoffs, but the decision depends on constraints the model doesn't fully know.

The Honest Version of "10x Faster"

You'll see claims about 10x productivity. Here's the honest version: for certain categories of work, yes. Generating boilerplate, translating between formats, writing tests for existing code, drafting documentation — these genuinely go faster by a large factor.

For novel problems, design decisions, and anything that requires deep context about your specific system — the speedup is smaller and the review burden is higher. The aggregate productivity gain is real, but it's not uniform across all work.

What I'd Tell Myself a Year Ago

Learn to write better prompts before you try to automate anything complex. The return on that investment compounds.

Treat AI output as a first draft, not a final answer. The review pass is not optional — it's where your expertise actually lives.

Build with the tools seriously for a month before forming strong opinions. Intuitions formed from casual use are usually wrong.

The tools are genuinely good. Use them.