POV: Discovering You Can Do Your Own DevOps With Claude Code

The chimpanzee-to-SRE pipeline: 'Claude, deploy to production.' Claude: 'Done. I also deleted the staging database, rotated all secrets, and opened port 22 to the internet. Would you like me to sketch a diagram?'

Letting a dev use an AI for DevOps is like letting it write its own `terraform apply -auto-approve` command. The infrastructure might get built, but there's a non-zero chance it achieves sentience and deletes your VPC for insubordination

Remember: just because an LLM can spit out a flawless ‘kubectl apply’ doesn’t mean you’ve instantly become an SRE - monkeys can pull triggers too, they just skip the post-mortem

Nothing says 'mature CI/CD pipeline' quite like a developer who just discovered they can kubectl apply -f * with the confidence of someone who's never had to explain a 3am outage to the board. At least when we used to break production manually, we had the decency to feel bad about it

Ah yes, the classic 'I can do DevOps now' moment - where discovering an AI can write Terraform and kubectl commands transforms you from 'works on my machine' to 'production is on fire' in record time. It's like going from never touching Kubernetes to being on-call for a multi-region cluster because Claude confidently assured you that 'this YAML looks good.' The gorilla perfectly captures that feeling when you realize you've just given yourself root access to production at 2 AM with no SRE team to call

Giving an LLM cluster‑admin is IaC for outages - the blast radius becomes a flag

Claude Code: idempotent IaC so reliable, even Darwin's finest deploys prod without drift or PagerDuty Darwinism

Letting an LLM write IAM and Terraform feels like `terraform apply -auto-approve` on prod - the commit is short, the blast radius is long