People keep asking what AI tools I use. So here's the honest answer — the actual AI stack for DevOps engineers that I reach for every day, with zero sponsorship and honest notes on what works, what's overhyped, and what I quietly stopped using.

I run infrastructure across 500+ VMs and 62 Azure subscriptions. AI is not a conference demo for me — it lives in my terminal, my editor, and my planning. Here's exactly what does what.

Cursor — the daily driver

Most of my day is Cursor. Terraform modules, Bicep, PowerShell. I write maybe the first 20 lines to show it the direction, and it fills in the rest. I still review every line, but reviewing is a different kind of tired than writing from scratch. Faster. Fewer stupid typos.

The numbers are real. Before, a Terraform module was 45 minutes of writing. Now it's 10 minutes of review. That's not marketing — that's my ordinary Monday.

Claude — the thinking partner

For the big stuff I use Claude. Designing a migration path. Figuring out why a monitoring query silently drops data. Planning how to split a deployment across banking change windows. I dump the full context in — constraints, regulations, team structure, what we already tried — and it's like brainstorming with someone who never gets tired and occasionally catches what I missed.

Last time, it flagged a cardinality problem that would have created 500,000 unique time-series. In production that would have been an ugly, expensive thing to unwind. Catching it before the deploy is exactly the kind of leverage I want from AI.

Debugging changed too. A gnarly Kubernetes issue used to mean two hours of googling. Now it's 30 minutes of guided hypotheses — I still make the calls, but I'm not alone with a search bar.

GitHub Copilot — the boring stuff

Copilot I use for the boring stuff. Tab-tab-tab. Variable names. Patterns I've written a hundred times. Nothing glamorous, and it doesn't need to be. It removes the friction from the parts of the job that were never the hard part.

ChatGPT — I stopped using it for infra

ChatGPT I stopped using for infrastructure. Too generic. It gives you something that looks right but doesn't know your naming conventions or your compliance rules. I burned an afternoon debugging output that was "technically correct" but didn't match our Terraform module patterns. For general writing it's fine. For infra with real constraints, it cost me more time than it saved.

The key insight — AI replaces typing, not thinking

This is the part that took me a while to land on: AI replaces typing, not thinking.

Documentation used to be something I dreaded. Now it gets generated from the code and I edit for clarity. Great. But the architecture? I still design it. The trade-offs? I still make the calls. The responsibility when production breaks at 2 a.m.? Still mine.

None of these tools will tell me that this particular team gets nervous about Monday deploys. None of them will decide a deployment should wait because the change window is too tight. None of them feel anything when production breaks. That judgment stays with me.

The catch nobody puts in the demo

The "5x more productive" headline is real. But it comes with a trap: AI makes you faster at everything, including making mistakes. Five times the speed with zero review is five times the damage. The engineer who uses AI thoughtfully beats both the purist who refuses to touch it and the blind adopter who ships whatever it hands them.

Here's the honest before-and-after of my week:

  • Terraform module: 45 minutes of writing → 10 minutes of review
  • Kubernetes debug session: 2 hours of googling → 30 minutes of guided hypotheses
  • Documentation: dreaded chore → generated from code, edited for clarity

I produce code faster now. Decisions take exactly the same amount of time they always did. Turns out those are completely different skills — and the second one is still the job.

So what should you actually run?

If you want a starting point: Cursor for day-to-day IaC, Claude for the hard reasoning and architecture, Copilot for autocomplete. Skip the generic chatbot for anything that touches your naming conventions or compliance rules. And test every tool on your real pipelines, not a toy repo — the only verdict that matters is the one from your own infrastructure.

Start today. Use AI. But keep your brain on.