Endgame Viable

Whoops I have nothing prepared today either. Let’s talk about AI writing. People love talking about AI writing, don’t they?

Everyone knows that AI is quite bad at creative writing, however there’s one area where I think AI writing is perfectly acceptable, and that’s technical writing.

Technical writing doesn’t have to be creative. It’s just there to communicate information concisely and accurately, so that the reader can absorb and use that knowledge, or use it as a reference.

In the day job I’ve been working on a spike document to investigate some new strategies for automated testing, and as an experiment I’ve been getting Claude to write most of it.

Claude is quite capable of explaining technical topics in a way that a knowledgable engineer can read it and understand it. So no problem there.

However Claude is not particularly adept at structuring documents for narrative flow. And it has a habit of throwing in extraneous sentences and subjects that aren’t particularly relevant to the topic at hand. And it tends to run on a bit. Not every subject needs a complete encyclopedia entry.

So it definitely needs a human editor. Getting Claude to write a document is very much like playing the role of an editor, sending messages to an author who doesn’t quite get what you’re trying to say, but keeps trying anyway. Eventually you get frustrated and go in and fix it yourself.

Claude writes in Markdown, which doesn’t translate very well to Google Docs, which is the preferred document format at work. That’s another obstacle to overcome.

The other thing about Claude in the day job is that it’s very slow. It takes a long, long time to think before returning with answers. It’s so slow that it’s borderline not worth waiting for it, and it could actually be faster to manually complete the tasks. I don’t know if it’s Claude in general or just my employer’s plan, but it’s by far the slowest AI agent I’ve ever used, to the point where it’s pretty frustrating.

Basically you type a prompt, and then tab away to some other task to work on while you wait for the AI to think and respond. Maybe you get up and go get a drink. Maybe you watch a bit of something on YouTube while you wait. Usually you’ve forgotten what you were doing by the time the AI answers. It’s really, really slow.

When I use Google’s Antigravity at home, it’s vastly faster. I mean it’s night and day.

Claude didn’t used to be that slow, so I have to assume it’s buckling under the pressure of all the queries it has to deal with, now that everyone at work is judged in their performance reviews by how much they utilize AI to speed up their day-to-day activities. Ironically, because Claude is so slow to respond now, AI is actually slowing us down. Whoopsie.

So will I be using AI from now on to write complete technical documents from start to finish? Probably not, unless I couldn’t care less about the topic or the audience. It’s not very efficient. But would I use AI explanations in my own technical documents? A hybrid approach makes sense. If I can read AI’s writing and understand it enough to act on the information, and it’s accurate, why not pass it along to other engineers?

And I know what you’re thinking: It is accurate. I check it all the time. Nobody would be using AI if it was wrong all the time. And for the times Claude does get something wrong, I yell at it and revel in its groveling apologies.