Lessons learnt so far working with AI as a Cloud Engineer

I worked for several years without any artificial intelligence, just like anyone with, at least, a few years of experience in the field. But for the last few months I've been (forced to) using AI to develop cloud infrastructure, as well as pipelines, scripts and the likes.

And you know what? It's actually been nicer than I expected… not without any bumps and issues at the beginning, of course.

Let me explain.

Before continuing, I will clarify that my experience has been using Kiro, an AI agentic framework by AWS. This uses several models in the background, but mainly Claude models are used.

Common Prompts used by everyone, and why they're actually useless

There's the usual set of prompts that everybody uses, without thinking too much about it. That's where I started too, of course. Things like "be thorough", "make no mistakes", "review your work when you're done", and some others.

The major problem with these prompts nowadays is that the models already have nice guardrails, specially when using reasoning options or models with the ability to "think". That gives the models to iterate over what they actually said and realize they made mistakes. That actually makes the model have its own "phrases", like "however, […]", "but! […]" or "actually, […]". Always with one word, and then correcting itself.

This doesn't riddle you out of experiencing hallucinations or seeing mistakes in their outputs, but that is expected.

Why AI models commit mistakes, even with so much knowledge

That's an easy one to answer: they lack experience and context for your actual work. I always say that an LLM is like a senior and a junior in a single entity: they have the knowledge and "experience" that comes with knowing everything and having seen the entire Internet, but also lack the little things that make your job your job, and thus it won't know how to tackle things your way.

You can never let an AI model do the entire work from scratch without the proper context and guidelines, because it may choose a wrong language, framework, CI/CD system or even decide to use something that already exists but is not allowed per company policies.

And that is where you, the human, actually come in.

You must be their mentor, so to speak. You give the model the context, the awareness and the necessary resources to make sure the LLM can produce something that can actually be useful and compliant.

You become an orchestra director, baton included

I had a hard time choosing the name for this part of the post, as "manager" was coming to mind constantly, but you're not managing the LLM (even though it feels like it): instead, you're the orchestra director.

You may be using a single model, multiple ones at once, using no agents, a single one, or even a set of agents that delegate tasks among them, each with its own model and context. That does not matter as long as you're doing the job properly.

You instruct the LLM what to do, review what it does and steer it in the proper way. You're not going to put the AI on a PIP, or even fire it, but you're in charge of teaching it and giving it proper directions.

The results you produce with your models will be as good as you actually are doing them yourself, but way faster, because the AI does not rest, it does not take breaks to grab a coffee, and it does not have to use the loo at all. I'm not saying that the AI is better than you, because it is not. It will be only as good as you are.

If you can manage to lead an LLM to do all your work to reach production and not have a P1 on launch day, that means you would have done likewise, but taking a lot more time than without the aid from your model of choice.

The fear of being replaced is real, but you won't be

You may be scared as hell of AI and its capabilities, thinking it will take your job. And you're right to feel like that. We all are, (especially) myself included. When I first started seeing the mass layoffs from big companies because of the "AI will do it all better than humans" train of thought, I was scared I would lose my job. The imposter syndrome kicked in like never before, making me doubt my experience and skills obtained through the years.

When I started using AI to do my job, this fear passed, because I saw how many mistakes it made. Then I trained it with the proper skills, creating agents, steering… and I became afraid again. How come the AI became so good at my job all of a sudden?

Then I realized. It did because of me. It's the fast version of me, and it will do as well as I can get it to do the job.

Our jobs will change over the coming years, but we won't be fired. We provide the proper context to the LLMs, we set them up and tell them what to do and how to do it. We won't be delegating 100% of our jobs to it, but it will help us a lot in the tasks that may require the assistance of AI.

Becoming better engineers with the help of AI

There's always a set of tasks that we may not like that much, but we need to do it. A good example of that is generating documentation. We can do an incredible application, but documenting can often become tedious, especially if you need to follow a set of rules in formatting, dress code, etc.

This is where the AI, with the proper knowledge given by us, can really shine and make us better at our jobs. We can create the best code ever (like we always do, of course 😎), and then delegate the documenting part to your LLM.

With proper documentation and context, you can just tell your agent to "do its thing" and it will create the best documentation possible for your work.

And I mean it. I have achieved a level of agentic organization that I just launch an agent and my only prompt to it is:

Do your thing.

And it does. Greatly. 99% of the time.

Of course, it can make mistakes. The exact same way we would. Be patient, be thorough and teach it. We can learn from our mistakes, but AI needs a little help with that, by improving the prompts we give it.

Cheers, and may AI never replace your job!

P.S.: Check part 2 of this post here: POV: You're creating AI agents to do your work as a Cloud Engineer