AI is changing how we work. The humans haven't changed much, though.

AI is everywhere. ChatGPT, Claude, Codex, GitHub Copilot, and probably another five(hundred) tools I haven’t even tried yet. Using AI for engineering work is quickly becoming less of a special thing and more of a default. And honestly: that’s pretty cool.

Need to understand an unfamiliar codebase? Ask AI.

Need to refactor something? Ask AI.

Stuck with a weird bug? Give it some context and see what it comes up with.

Want to brainstorm an implementation or get a first draft of something? Also AI.

These are just a few tiny examples - there are plenty of things these tools are genuinely good at. They can make onboarding to a large codebase easier, speed up tedious work, help with debugging, give you different perspectives on a problem, and simply get you unstuck.

Of course, there are also the usual buts. Shit in, shit out. AI can confidently produce nonsense. There are security implications. It’s easy to accept an answer without really understanding it. And if we stop paying attention, we can produce a lot of very polished-looking AI slop. But there is another aspect I’ve been thinking about lately a lot: The person on the other end.

We’ve seen this before

In a way, this reminds me of the Agile years. When Agile became a thing, there were lots of genuinely useful ideas around it: cross-functional teams, short feedback loops, regular reviews, retrospectives, and actually talking to the people involved instead of throwing things over a wall.

I like working that way. I still do. But some of it always made me wonder: Wait. We needed a framework to tell us that different people working on the same thing should talk to each other? We needed a meeting to realise that maybe the way we’re working isn’t working?

Obviously, Agile is much more than that. But some of the things that were presented as revolutionary also sounded a lot like… common sense. And I think AI has a similar trap.

We have all these new and incredibly powerful ways of producing information. But the person receiving that information hasn’t suddenly become more efficient at reading it.

AI makes it very easy to produce a lot of information

Imagine a developer reporting a problem to the infrastructure team. The ticket contains a beautifully structured explanation of the problem, several paragraphs of analysis, possible causes and a proposed solution. Great. Except nobody actually knows whether any of it is correct - the developer just pasted the output of an AI tool into the ticket without really checking it. Meant well, just wanting to help - but not really helpful.

Or a client sends a bug report to a service provider, including a suggested fix. The infrastructure engineer reads the report, recognises the issue, agrees with the proposed fix and applies the changes. They report back:

Fixed. Applied X, Y and Z as described in the ticket.

Client:

What exactly did you change?

Infra:

As described in the ticket. X, Y and Z.

Client:

Ah. I didn’t actually read that, just pasted what Claude wrote.

And then there’s the other direction. The service provider sends a detailed fixing report to the client. It’s technically correct, very thorough and several pages long. Somewhere in the middle of all the AI-generated explanation are two things the client actually needs to do. Good luck finding them.

More output doesn’t automatically mean better communication

I think this is where we need to be careful. AI is incredibly good at producing information. It can turn a five-minute conversation into a two-page summary. It can analyse a problem and give you ten possible explanations. It can turn a handful of notes into a beautifully structured report.

But producing information and communicating effectively are not quite the same thing. The person on the other end still has to read it. They still have to understand it. They still have to figure out what’s relevant to them. And they still have to decide what to do next.

AI doesn’t remove any of that. If anything, it makes it easier for us to overlook. When producing another 1,000 words costs basically nothing, it’s worth asking which 100 words the other person actually needs.

So, do we need an “AI Knigge”?

Not another framework. Not another process. Just remembering some basic rules for using these tools without forgetting that there are humans involved.

Keep it simple, stupid.

KISS still applies. If the person receiving your message needs three things from you, give them the three things. Don’t send them the entire AI conversation that helped you arrive there.

Understand what you’re sending.

Using AI to analyse something is fine. Pasting the result without understanding it is a different thing. If you’re asking somebody else to act on it, you should at least be able to answer: Do I actually believe this?

AI output is a draft, not a source of truth.

Check the commands. Check the assumptions. Check the security implications. Check whether the proposed solution actually fits your environment. Especially when you’re asking somebody else to execute it.

Communicate for the person receiving the message.

The fact that your AI tool needed 2,000 words to explain something doesn’t mean your colleague needs 2,000 words. So: Summarise. Highlight the important bits. Put the action items somewhere people can actually see them.

Don’t outsource responsibility.

“I just pasted what Claude gave me” isn’t really an explanation. AI can help us brainstorm, write and communicate. It shouldn’t become an excuse for not doing those things ourselves.

The tools change. The people don’t.

I don’t want to go back to a world without AI-assisted engineering. Quite the opposite. I’m really curious about where this is going, and I’m already seeing how useful these tools can be in everyday engineering work.

But the better our tools become at producing code, analysis and text, the more important it might become to remember the things they don’t replace:

Context.

Judgement.

Curiosity.

Communication.

And responsibility.

Because at the end of the day, the person reading your ticket or email or whatever doesn’t care whether Claude, Codex or ChatGPT produced the first draft. They just need to know: What do you need from me?

header image created by Canva AI