How to use LLMs to code critical software for you

LLMs are getting much better. We soon won’t be needing programmers, yay! We all know that software engineering consists of writing “for” loops all day and not much else, and I am glad that I won’t be writing “for” loops anymore. Now I can move on to things that really matter, like scheduling meetings and strategy sessions to discuss how to manage LLMs. Here are some tips that will help you use LLMs to write code for the important software in your company, even if you don’t have any coding experience.

Is this for real? (spoiler) >

Stick with the models you know

ChatGPT 3 is pretty good. There’s no need to try Claude 3.7 or ChatGPT-4.5. I know that there are newer models coming out, but who has time to try them all. If you have experience with a model, don’t try any new ones. The pace of progress has been kind of slow in AI, so I suspect newer models are about the same. Plus, the progress in AI such as LLMs is about to hit a wall.

Don’t test your code

Once Grok3 -for example- has generated your code, run it directly on Production. Real databases, real data. This is the only way to test that it works properly. I know your technical team always insisted on testing things a million times in different environments, but that has always proven a waste of time, since the software didn’t have so many terrible bugs, as far as you know. Or at least, not bugs that your users have complained about.

Testing wastes your valuable time. Besides, you would need technical people to automate those tests. Here’s a better idea: have your LLM of choice write tests for you. It doesn’t matter if you don’t understand what the tests are, as long as they are all passing. You will know the tests are passing because the letters are green, and probably you will see check-mark next to them. If, instead, you see red “x”, it means that they are not passing, but guess what? You can have the LLMs fix that for you too! What the test is actually testing is the least important thing.

Don’t check your costs

All models, even newer experimental ones, are darn cheap. Just ask them about everything, even when you suspect that it won’t work. Leave them running in agent mode and ask them to fix anything and everything inside your company. Deployment pipelines, architecture, payroll policy, you name it. Every once in a while (but not too often) check what they have been up to.

And don’t even think about checking your LLM billing. Think of it this way: you are automating your whole technical team, so even if it’s > $X0,000 a month, you are still saving a lot of money. Even better, when AGI comes, your can automate your non-technical staff too.

Use this prompt: fix everything, everywhere, all at once

Your engineering team used to tell you to use small and specific prompts with measurable side effects, and that’s why they had to be let go. You don’t have to play that game. LLMs are quite capable of understanding complex and diverse instructions in the same prompt and fix everything in your codebase. Also, if the LLMs doesn’t fix it right the first time, you can ask it to fix it again all the times you want.

Don’t get any engineers involved

This is very important. Even if you do all the above, getting good engineers to look at the code will get in your way. They will start complaining about “tech debt”, “performance”, and “reliability”. They haven’t caught on: LLMs can fix all of that if you let them.

If you need someone to look under the hood and read the code the LLMs generated (yuck! who would want to do that?), think about other departments in your company, like HR, Design, and Product. After all, most programming languages are almost human-readable. And surely, anyone can deal with code, it’s not that complex. How the code works should be pretty obvious just by looking at it.

In any case, the code being written doesn’t matter too much, it only matters what your product does. That these are kind of related is a horrible accident that you will have to fix by asking the LLM to make the code behave in such a way to give you the product you want. But don’t worry, I promise you that this won’t make you a “programmer”, you are only designing how a deterministic digital system will behave.

Follow every piece of news about AI

It’s important to stay informed. Soon, we won’t need programmers, and AGI is around the corner. I wouldn’t know this if I didn’t follow closely all the news in the AI area. Read anything you can get your hand on about to Anthropic, OpenAI, Mistral, Grok, DeepSeek, and any company who says that they have a new model around the corner that will obliterate the need to work. As you should already know, in the not-so-distant future, nobody will need to work; the AI will do everything for us.

---

OK, now for real

Writing code is only a small part Software Engineering. If LLMs can help relieve programmers from the burden of programming trivia and syntax (a big “if”), they can focus on the important stuff: system design, the quixotic art of dealing with chaotic complexity in an useful way.

Even with LLMs writing almost all the code, companies with competent engineering teams that steward the software to success will beat other who don’t. It’s not about who writes more code faster.

I won’t abuse satire going forward.