Raymond Brigleb

I came across How To Write With an LLM the other day and I thought it was a nice piece with some good advice.

Sure, it’s brief. It only has two points! But I agree wholeheartedly, and they’re worth repeating.

The first rule is “You may not use a single word an LLM suggests to you” and it kind of sounds like a joke. But it’s not. I use these large language models a lot, and they’re really good at things like code. For prose, however, I can smell them a mile away. Several times this year I’ve gotten things from clients that I like where it’s absolutely clear they used an LLM to help them write it.

The problem is that if I can tell an LLM did the writing for them, who’s to say it didn’t do the thinking as well? And if the LLM did the thinking, how can I trust what I’m reading? I don’t know for sure the client actually knows what they’ve promised, or what they’re asking.

I use straight LLM output quite a lot for code, but somehow this is a bit different. I do ask the system to provide proof that things work, to standards we’ve already agreed upon. And I understand the technologies, so I know if it’s working or not. It’s hard to enforce this kind of rigor on the written word, in prose!

The second rule, also important, is to “avoid encouragement.” The point here is that the LLM tends to glaze you, and will constantly encourage your mediocre writing. But writing is hard, good writing involves a lot of thinking, and like with my point above, you cannot let the LLM take that step away from you. It will constantly suggest that your writing is excellent, but you have to trust your gut: no, not yet! It needs time to get good. Give it that time.

I’ve been thinking a lot about using AI in my work lately… in particular, LLM-based AI in an agentic environment. I’m frequently surprised at their rate of progress, and what they can accomplish given the right inputs. But I also think a lot of folks are going in with the wrong expectations.

It’s natural to go in thinking you should be asking for what you want. For example, maybe you want a new logo for your business. So you might say what your business is, give a brief description, and ask for a logo. But you’re never going to get a good logo that way, you’re going to get a very generic output.

This type of output is not acceptable. Sure, do things like this for a living, but that doesn’t mean I feel differently about a song or painting or other creative work generated by this kind of AI. They all generally suck. If you need a website, you’re still a lot better off picking a quality template and building from there, rather than asking an AI to build a website for you. It’s just not going to be good enough.

However, I often say that the best way to get a good X is to have a lot of X. Photography is a good example of this. I’m not a great photographer, but I’m a pretty good one. And I get good results by, more than any other technique, just taking lots of pictures and throwing most of them away! I encourage you to do the same.

This is where an LLM-based AI is most useful. You can ask it to come up with a lot of ideas, very diverse ideas, based on a lot of good context about a project. None of those ideas is good enough… but I can often see a lot of ideas in those designs I might not have had myself. The AI doesn’t get bored, doesn’t get tired, and doesn’t get tired of your shit. You can have it build you 100 different home pages if you want, and pick and choose the best parts.

So that’s my advice. Don’t try to get your AI to have good ideas for you. Get it to have a lot of bad ideas, so you can skip to the good part.