Ryan M. Patrick

Well this is a hot topic right now…

Last post, I showed why AI is a bad drafter. But there’s a lot that an LLM can do, and it’s not all bad – just don’t have it write any words for you!

I’ll break this post into two categories – what I’ve found that AI is good at, and what it is most definitely not.

What AI is good for

Calculations

If you give it the right inputs – and can do a quick sanity check – an LLM can save you a ton of time if you’re doing a complex series of math equations.

For example, in The Saturn Anomaly, I have the space shuttle Atlantis doing a pair of complex orbit changes. One from its notional launch trajectory to a polar orbit to retrieve the Black Knight satellite, the other to put it on a course that would let it reenter and land at Dreamland/Area 51’s dry Groom Lake.

I’m an aerospace engineer, and the shuttle’s performance characteristics are pretty well-known. The math isn’t that hard, and there’s tools like STK that make it trivial. But I can also feed it into Grok or Claude and get the exact orbital parameters and apogee and perigee. I would not recommend this for everyone – there’s a sanity check needed that unless you have an engineering degree or a space background you’re not going to be able to do yourself – but it saved me a ton of time when I was working on that book.

Brainstorming

This is probably the most controversial one here. Some people are against LLMs for any and all reasons, some are more against it for creative purposes (I lean in this direction), and other writers/creative types are using it for everything. Discussing story ideas with an LLM and brainstorming together is a hot topic to say the least.

But LLMs/AI are a tool. They give you an output based on your input – if your input is bad, you’re going to get a bad result (GIGO!). And I’ve tested this:

I fed Claude the first two Grace Parkowski novels (they already have Lag Delay, thanks pirates & Anthropic) and the outline for #3, as well a high-level idea of #4 (I don’t like my outline, long story). And I asked it for some sequel ideas. And boy did I get some trash.

I gave it my full stories and still it didn’t understand them…because it can’t! It’s not thinking! It’s linear alegbra trying to predict the next token.

After some back and forth, it finally understood what I was trying to do, but still can’t produce anything useful.

Until I – the creative, not the LLM – started feeding it my ideas:

And a little later, after a lot of my own ideas – again, not the LLM – I came up with

And that’s what I needed – I’ll go knock out the outline myself based on that. That’s probably Book 5.

Research

Google sucks. It’s well-known at this point – its parent company Alphabet is an ad company and AI company more than it does search. But man can AI (particularly Claude and Grok) fill that gap. For example:

Again, a huge time saver, and everything is backed up with links. But I think most people are already using LLMs for this so I’ll keep going.

Consistency & Typos

Another more dangerous one. Some people are super-against their manuscripts getting fed into LLMs – and I get it. But they’re going to get your stuff one way or another. And people want to read human works – see my last post – AI writing sucks, and I honestly don’t think that problem will ever be solved. People want to read human-written works with soul and no LLM is ever going to produce that.

But you can feed a manuscript into an LLM and ask it for help with a specific task. The two that I use are consistency and proofreading/typos. It’s ok on the last one – you’ll still want to to a pass yourself – but the former is much more interesting.

In Crush Depth, there’s a lot of numbers – depths, distances, years, pressures, etc – that I had to keep track of. And, again, I’m an engineer – I’m good with this stuff. But I’m not perfect. Feeding my manuscript to an LLM with specific instructions “check depths, distances, etc” produced an incredible report showing where I had messed up.

Once again, it’s not doing anything creative for me, but it is leaning into the LLM’s strengths. Garbage in, garbage out.0

What AI is bad for

Names

I suck at character names. Always have. And I thought LLMs would be able to help me with this.

But they are terrible.

This article summarizes it better than I could, but there’s a small default set of names that LLMs always fall back to. “Lena Voss,” “Elena Voss,” and others will show up no matter how much you try and prompt engineer yourself out of it.

Even specific instructions have failed for me – I’m outlining a WWI supernatural/fantasy novel, and I want a Polish-American name. Should be pretty easy, but it keeps giving me last names that are just not Polish. Back up to typos – I have a female NCO in Crush Depth with an African name (Okonkwo), and my typo runs on the manuscript kept spitting back that I needed to change it…and it was inconsistent each time as to what I should change it to!

Run, don’t walk, away from LLMs having anything to do with your character names.

Anything Creative

I covered this a lot in the “Brainstorming” session above, but LLMs are awful for creativity. This is a feature, not a bug – they have limited training data and will default back on standard, strict rules to align with their model. But it’s just so laughably bad at the route it takes and the fallbacks it keeps coming back to.

After the Stonehenge idea, I tried to use the same Claude conversation to try to come up with another idea – I love the Nazca Lines in South America, love the mystery there.

But Claude couldn’t help me out at all here. And this is my fault. Unlike with the Stonehenge plot idea, I didn’t have enough to go on, even when I kept throwing ideas at it. It just can’t do anything creative and reverts back to the tropiest of tropes if you don’t have a strong creative vision.

…yeah, it’s stuck on counterdrug stuff, even though I didn’t want it to.

Moral of the story: LLMs are a creative tool. They’ll amplify your creativity but they’re like gain in an electronics system. If there’s no signal to boost, anything times 0 is still 0.

Developmental Feedback

Last one – I have tried multiple times to build a dev editor or beta reader using an LLM. I’ve tried Gemini, Sonnet, Opus, and Grok.

All of them are awful.

This comes back to the creativity issue. The LLM is predicting the next token, nothing more. It can’t take a step back and diagnose the why, it only knows what’s in its model and its training material. You can feed your manuscript into an LLM, sure, and maybe there’s some use for it. “Is Chapter 11 too long,” “do I have too much dialog in Chapter 23,” etc, but you need to give it specific instructions like I talked about in Consistency/Typos above. It’s not going to diagnose huge pacing issues or an unlikable character on a whim. For that, you need beta readers or a real editor.

Last Note

This is how I use LLMs. Like I put at the beginning of my last post, I am very pro-AI/LLM, I use it every day at my day job (have Claude Code doing a huge refactor for me while Codex is troubleshooting a networking issue as we speak). I think there’s usages for it in my writing, but obviously with huge limitations. If you don’t have a clear vision and creative story ideas, an LLM isn’t going to help you one bit. I’m not saying AIs are right or wrong, and feel free to hate me after reading this – I’m just presenting how I’m using them and what I’ve found useful in case it helps someone out. I’ll do one more post in this series on the different models and what they are good/bad for in my experience.