5 Things I Learned About AI-Assisted Engineering This Week (and 1 That Backfired)
It was a week of experimenting with AI in real workflows — not demos, not toy projects, actual production code. Here's what stuck. 1. The prompt pattern matters more than the model I spent most of the week validating this. Swapping models with a mediocre prompt gives mediocre results. A tight, structured prompt on a weaker model often beats a lazy prompt on a frontier one. The framing is the work. 2. Refactors are the best use case nobody talks about Everyone focuses on greenfield generation. But handing AI a gnarly legacy function with a clear "here's what it does, here's what it should do, here's the constraint" prompt is where you get the real time savings. A task that looks like a 3-day slog can collapse into hours if you nail the context window. 3. AI doesn't replace code review — it changes what you're reviewing for You stop catching typos and start catching logic. That's actually a better use of a senior engineer's brain. Let the machine handle the syntactic noise. 4. Piping AI into your existing CLI tools is underrated Not everything needs a chat interface. Wrapping a prompt pattern into a shell script or a Makefile target means your whole team gets the benefit without changing their workflow. Low friction = high adoption. 5. ❌ The one that backfired: using AI to write tests first I tried feeding AI a feature spec and asking it to generate tests before the implementation. The tests were coherent but subtly wrong — they tested the assumed behavior, not the correct behavior. It created a false sense of coverage. Writing the implementation first, then using AI to expand test cases, worked much better. The throughline this week: AI in engineering is a workflow design problem, not a tool-selection problem. How you structure the interaction — the order, the constraints, the context — determines whether you get a 10x or a 0.5x. If you want the specific prompt patterns behind the refactor approach I kept referencing, I've packaged them (along with the workflow structure) into a concise playbook: grab it here. More next week. Stay concrete out there. 🛠️
