I’m a DevOps engineer (about 15 years) and in a previous life was a software engineer (15 years before that). My employer is pushing hard on AI so I reluctantly started using Claude at times. I must say that I’m fairly impressed when it comes to relatively easy tasks. We’re a large AWS user and have developed a fairly complex in-house set of python tools that encapsulate things like Terraform and Ansible. We have about 15 or so AWS sub-accounts that span logical groupings, so our IAM configuration alone was fairly complex.
I was able to point Claude at our IAM configuration and tell it to create a set of policies/roles to allow a host in one environment to access resources in a read-only manner across all our accounts. Since I’m not an IAM expert it would have taken me a few hours to figure out what it did in under 10 minutes. Two of my team reviewed the proposed changes and were perfectly fine with them.
I’ve also had it write python scripts that do things like call AWS APIs, collect JSON results, and compare it to contents pulled from a git repo of configuration data.
For relatively simple tasks like these it can be a time saver. But you still need to sanity check everything it does. I’ve seen it skip steps (like not applying IAM policies to all our accounts), and when you point it out it will apologize and fix things. But it’s that sort of failure that makes me still be wary of AI. Like why only update a subset of things and fix it only after I point it out? “All” means “all”, not “some”…
For more complex things I’m still very reluctant to trust it. When it comes to that I may use Claude to encapsulate a few API calls, but then I’ll rely on my own expertise to add in all the really important logic.
Same take here. I usually break things down to simpler tasks first and it does better. But it tends to get lost fast if things get too long or too complicated
I’m a DevOps engineer (about 15 years) and in a previous life was a software engineer (15 years before that). My employer is pushing hard on AI so I reluctantly started using Claude at times. I must say that I’m fairly impressed when it comes to relatively easy tasks. We’re a large AWS user and have developed a fairly complex in-house set of python tools that encapsulate things like Terraform and Ansible. We have about 15 or so AWS sub-accounts that span logical groupings, so our IAM configuration alone was fairly complex.
I was able to point Claude at our IAM configuration and tell it to create a set of policies/roles to allow a host in one environment to access resources in a read-only manner across all our accounts. Since I’m not an IAM expert it would have taken me a few hours to figure out what it did in under 10 minutes. Two of my team reviewed the proposed changes and were perfectly fine with them.
I’ve also had it write python scripts that do things like call AWS APIs, collect JSON results, and compare it to contents pulled from a git repo of configuration data.
For relatively simple tasks like these it can be a time saver. But you still need to sanity check everything it does. I’ve seen it skip steps (like not applying IAM policies to all our accounts), and when you point it out it will apologize and fix things. But it’s that sort of failure that makes me still be wary of AI. Like why only update a subset of things and fix it only after I point it out? “All” means “all”, not “some”…
For more complex things I’m still very reluctant to trust it. When it comes to that I may use Claude to encapsulate a few API calls, but then I’ll rely on my own expertise to add in all the really important logic.
Same take here. I usually break things down to simpler tasks first and it does better. But it tends to get lost fast if things get too long or too complicated
Well written. This is pretty much exactly how our dev team is using LLMs. Verify everything, but it sure does save time.