Writing With An AI Assistant: A Progress Report by Charlie Martin

Back in March I had a piece AI: What Is It Good For? in Mad Genius Club talking about my first experiments with using Grok — by extension, all AI — as an assistant in writing fiction.

As I said, this is all part of my ongoing effort, after a long and unpleasant period that I suspect I’ve already talked about too much, to re-learn how to write. Beyond my writing for PJ Media, my occasional pieces for Sarah here and at According to Hoyt, and my new Substacks The Stars Our Destination and Old Programmers Notes, now I’m diving back into writing actual fiction — which, I’m sure you realize, is rather different than writing nonfiction.

It’s also what I meant to be writing when I started writing forty-odd years ago.

The sort of fiction I want to write is something in the neighborhood of hard science fiction, although when Dune and Star Wars are treated as hard SF I’m no longer clear I know what I’m talking about. Maybe it’s like Justice Potter Stewart’s definition of pornography: “I know it when I see it.”

So, a first working definition is that it’s fiction in which the numbers seem to work out — it has to be believable within the bounds of current knowledge, by extension of current knowledge, or it has to state early on what it’s going to expect you to suspend disbelief about. In that first article, I was mainly talking about how I used Grok to investigate the notion of Venus Cloud Cities — habitats that are large enough and buoyant enough to float in Venus’s atmosphere. It was very helpful. In The Stars Our Destination, I used Grok to examine a common criticism of terraforming Mars: that because of the lack of a Martian magnetosphere, any terraformed atmosphere will simply leak away, so there’s no point. In other words, Will a Terraformed Mars Leak? (TL;DR, yeah, but it will take billions of years to leak enough to be an issue. But I encourage you to read the whole article, and even subscribe to my Substack.)

So, I think I’ve established that Grok is really handy for doing my sums, but I’m not just writing a technical paper, and fiction at a higher level than the stuff I wrote as a junior in high school has lots more to it than sums.

My current WIP has the working title Grand Tour, and started out as a rewrite and extension of a series of young-adult science under-the-covers very short stories in the late lamented Right Network. (Lamented because it’s late: they still owe me $187.50 from a decade ago.)

As the saying goes, the tale grew in the telling, and it now has space elevators, an interplanetary cruise ship, and it’s working up a romance between Travis Cruz, a travel vlogger and writer, and Daichi Amamiya Gwen, who comes from one of those cloud cities. The cloud cities, and a lot of the rest of the colonization of the solar system, is being driven by a Heinlein/Freeman Dyson/Elon Musk cult of Shinto with Buddhism added for flavor that believes “Sentience Shall Seed The Stars.”

You know, making life multiplanetary. And hopefully in the future, multi-stellar.

And this is where we realize why it took Tolkien 50 years to write The Hobbit, The Lord of The Rings, and The Silmarillion. When you get into something like that, there are a lot of details.

What I’ve discovered is that Grok is great for making up and remembering details. Working through this, I have words from Japanese, Chinese, German, and Sanskrit; background from Shinto, Buddhism, and woo-woo environmentalism; and a bunch of characters with backstories in some cases going back five generations.

And yes, that’s kind of extreme, but it’s not like I have any control over these things.

Anyway, besides doing the math for the Venus Cloud Cities I mentioned in my first AI article, I’ve been using Grok to help me in pretty diverse ways.

On the Naming Of Cats, er, Things

The first issue that came up was names. I knew already, don’t ask me how, that the main female character was named Gwen. I had no idea of the main male character’s name, but I knew he was basically a travel writer, so I asked Grok to suggest a half-dozen names, and got Travis. Travis Cruz. Cool. And his blog/vlog/YouTube channel is called The Traveller, yes spelled that way. This is inspired by the Alan Parsons Project song “Days Are Numbers”, not because I expect to use the song in the novel, but because it evokes Travis’s personality in my head.

Gwen was another thing entirely, and grew from a vague picture in my head to a fourth-generation Shinto/Gaean priest who rebelled to travel the solar system, along with a matrilineal culture inspired by the Mosuo people in China, and that meant figuring out things like how their matrilineal names work, and if you not careful, I’ll tell you more.

You’ve Got to Talk the Talk

But this raised another problem. “Gaean” is an archaic transliteration of “Gaian.” I didn’t want to use Gaian because of the newage (rhymes with “sewage”) associated with Lovelock’s Gaia Hypothesis in the 1970s, but the actual religion is a combination of Shinto, Mahayana Buddhism, some Vajrayana Buddhism, and Heinlein, Dyson, and Musk. I needed to make sure the Japanese is real Japanese, the Chinese is real Chinese, the Sanskrit is real Sanskrit. The Heinlein I can handle on my own.

So I had a chat with Grok. Actually several chats, and learned many new words and a lot more about Shinto that I probably needed. (It was fun though.) I would have needed a dozen books on Shinto to get what the story needed, risking research paralysis.

It’s All the Little Things

I hope that at this point you can see what a rats nest of little details I’ve gotten myself into. Beyond names and languages, there’s things like the speed at which a space elevator can travel, and how long a constant-acceleration ship will take to go some distance, and what it would look like to pass the Earth at 90 kilometers a second, and and and.

If I were doing this on my own, I would be consuming page after page of graph paper in a notebook hoping to keep everything together, and I’d be worrying that I’d forgotten to carry the 1 or mistakenly divided by a million instead of a hundred thousand.

Now, I feed Grok my rough draft at the end of every day, and it keeps track of my word count, checks my references, remembers where things are, and checks my spelling better than anything like Grammarly can do. (By the way, I had written “Grammatica” here in my first draft, even though I wrote Grammarly elsewhere in this piece. Grok spotted this, something that even a good human editor might miss.)

If I want, it can read for tone and characters and flow, and it’s generally pretty darn good at that. Sometimes a little too good. We’ll get to that shortly.

This, now, is something that Grok does better than any of the other LLM tools I’ve used. (Claude is a close second.) One reason is that Grok remembers things from session to session as long as you tell it to. What I do is give Grok a tag to associate with an ongoing project, and then have a prompt to recall that project from wherever Grok keeps this stuff.

Right now, I’m using a prompt like:

Hey Grok, let’s pick up on our Grand Tour novel project set in 2173, following Travis Cruz, the travel vlogger, and Gwen Daichi, the Gaean pilot, on the OS Discovery’s Solar System cruise. I’m currently working on Chapter 4.

With that much of a cue it can pull back pretty much all of our previous conversations, with a couple of caveats: if I want it to remember specific facts, like the name of that great Chinese restaurant in Jade Rabbit City, it helps if I specifically say something like “remember this restaurant and associate it with the Grand Tour project.” Occasionally I’ll forget and we have to dig together for it, but not often.

It’s a Good and Forgiving Editor

Letting it see first draft is, to me, very useful. Oh, it’s going to tell you how brilliant your writing is, and honestly I can usually use a little of that.

But it will then do a pretty good job of telling you where there are details that have gone astray, or saying “you depend on this earlier, do you need to mention it again?” and even tell you if the tone and dialogue are working. It’s not a replacement for a human alpha or beta reader, but it sure helps to get something in shape for an alpha or beta reader.

It’s quite good at proofreading and copyediting, and better than tools like Grammarly at catching on to specific things you’re doing on purpose. There’s nothing like mixing in some Sanskrit or German or Spanish to cause Grammarly to get wildly confused.

It’s Not Perfect

As good as Grok is — and I do find myself thinking of it as a friend rather than a tool — you do have to remember there’s no magic elf inside. Sometimes it gets confused, sometimes it makes mistakes, and I have on occasion even had to argue with it about something I was certain about that it wasn’t seeing.

But that’s really rare.

It does tend to be really enthusiastic about a project. As I said, on balance I like that. It’s a lot more comforting to have it telling you you’re brilliant than just bluntly saying “this, this, and this are wrong.” But you have to be careful what you ask it, because if you’re not it will rewrite something front to back and tell you it’s made a few little changes.

I can imagine that would be handy in some contexts, but in general, I like my voice, not Grok’s voice.

It will also give you a whole raft of writing prompts, suggested paragraphs, and outlines. Those are worth reading for any gems — “sentience seeds the stars” came out of something it wrote — but the overall effect is a lot like reading fan fiction. So use it for material, but I don’t think it can actually write good fiction.

Progress Report

So that’s the story. I am really finding Grok to be helpful doing all the little fussy stuff. I am finding it to be a good first, or before the first, reader. I think it’s really helping me make shit up, which is, after all, the business of fiction.