How AI tools actually fit a working engineer's day
Cursor, Claude Code and friends don't replace engineering judgment — they move it earlier in the process. Here's what changed in mine.
I use AI coding tools every working day — Cursor and Claude Code mostly — and the honest summary is: they did not make typing faster, they made deciding faster. The bottleneck in real product work was never keystrokes.
What actually got faster
- Exploration. "Where is the tenant resolved in this codebase?" used to be twenty minutes of grep. Now it is one question, and I spend the saved time reading the two files that matter.
- Boilerplate with judgment. DTOs, guards, migrations, test scaffolds — the shape is known, the details are project-specific. AI does the shape; I review the details.
- The second opinion. Before a refactor lands, I ask the tool to argue against it. It finds edge cases I was too invested to see.
What did not change
Architecture is still a human job. An AI will happily generate a design that works in the demo and falls over at the first multi-tenant edge case, because it optimises for plausibility, not for your production traffic. Code review discipline matters more now, not less — generated code arrives with confidence and needs the same scepticism as a junior engineer's first PR.
The workflow that stuck
- 01Write the intent in plain language first — a short spec, even three sentences.
- 02Let the tool draft, inside the real repo, with the real conventions visible.
- 03Review the diff like it came from a stranger.
- 04Keep the tests yours. If you can't explain why a test passes, you don't own the feature.
The engineers who benefit most from these tools are the ones who already knew what good looks like. AI amplifies taste; it does not substitute for it.
Enjoyed this one?
I write about what I build. Let's talk about yours.