Most cloud engineers are still copy-pasting CLI commands from ChatGPT into a terminal. You ask how to list running EC2 instances, you get a command, you run it, you paste the output back and ask the next thing. You are the middleware between the AI and your infrastructure. I made a short course showing what happens when you stop doing that and point an agent at the infrastructure directly. This post is the gist of it, and the clips are embedded so you can watch the parts you want.
From a broken server to a fixed one in under a minute
The first clip starts with an EC2 web server behind Apache that goes unreachable after someone tightened a security group, and nobody knows what changed. I give the agent one sentence: our web server is unreachable, it worked this morning, investigate why traffic can't reach it. It hits a "must specify a region" error, self-corrects to US East 1, walks the CloudFormation stacks, the security groups, then the network ACLs, and comes back with the two blocking layers and a fix. I say run it, it deploys, the server responds. Doing that hunt by hand is usually 20 to 30 minutes in the console.
One prompt, one production VPC
The next video is about building, not repairing. From a single prompt Claude Code writes Terraform for a foundational VPC: a 10.0.0.0/16 CIDR, two public and two private subnets, an internet gateway, a security group, and VPC flow logs to CloudWatch, with variables for region and CIDR and no NAT gateway so I don't pay for one. It splits main.tf, variables.tf, and outputs.tf on its own, deploys all 23 resources with my existing CLI profile, and I launch a tiny EC2 into the VPC to confirm it serves traffic. Then I ask it for a Mermaid diagram of what it built, and it draws one.
Deploy, manage, tear down, zero CLI typed
The pilot lab runs end to end. Claude Code generates a CloudFormation template from plain English, deploys three EC2 instances plus a Lambda and API Gateway, stops only the instances tagged dev, tests the greet API which returns "Hello Cloud Yeti," tags every resource with managed-by Cloud Code in a bulk audit, then deletes the stack and confirms nothing is left running. I never type an AWS CLI command by hand the whole way through.
My take
None of this removes the engineer. I still describe what I need and check the plan before it runs, and my cloud knowledge is what lets me catch the agent picking a weaker approach and redirect it. As Karpathy put it, the hottest new programming language is English. What changes is where the time goes, away from typing commands and toward deciding whether the plan is right.
Work with me
I specialize in helping cloud teams point AI agents at real AWS infrastructure safely instead of at demo toys. If you want help wiring Claude Code into your own setup with the right review gates, book a call at cloudyeti.io/meet.
Book a call