My previous post discussed some of my experiences with LLM-assisted creative writing, and the basics of prompting for stories if you want the LLM to write realistic characters rather than falling into stereotypes. This one talks more about different ways in which you can actually use LLMs as co-writers, rather than as things you prompt.

I cover revising old content, spicing up dialogue with additional description, acting as a literal co-writer, and just having a fun and slightly deranged writer persona to discuss the story with and brainstorm. You'll want to do some rather heavy editing if you want LLM-produced prose to be actually good, but it can be a very good tool for helping get unstuck and producing first drafts.

Editing bad responses

I’ll be talking a lot about editing the LLM’s responses. One obvious question is, how do you do that, given that the web interfaces for Claude and ChatGPT don’t offer an option for it?

Back when GPT-4 was the best creative writing option around, I would do a lot of co-writing in the OpenAI Playground. This is a great option in that if the LLM produces a response that’s kind of good but you don’t like some of it, you can just edit it directly. The equivalent option for Claude is the Anthropic Workbench. (There’s also character.ai , which allows for editing the bot messages, but their model isn’t anywhere as good.)

However, both the Playground and the Workbench are billed according to their respective APIs rather than using one’s monthly ChatGPT/Claude subscription. This can get expensive fast - on occasion, I've ended up spending something like $15 a day on API costs. 

If you’re willing to pay those costs, there’s also other software such as SillyTavern that offers LLM creative writing tools. I’ve personally preferred to save the money, so I haven’t familiarized myself with those.

Fortunately, there are some easy tricks that you can use to edit LLM responses while using a chat interface.

The most straightforward is just manually copying the whole story so far into one text file, making any edits to that, and then starting a new conversation where you paste the file as the first message.

The second is to take a response that the bot generated and edit your own earlier message so that instead of the prompt, it has the bot’s response (with any edits you want to include) and your next prompt.

So for example, if we previously had

Kaj: Write a scene of a teenager having breakfast with his parents. He has been learning differential equations recently and is eager to explain those to his father, who is mathematically interested but never studied it past high school.

Claude: [scene with Marcus explaining differential equations]

Kaj: Great. Now write a follow-up scene where Marcus is anxious about whether a girl at school will like him or not. He thinks that she's both pretty and kind, and also into environmental activism, which Marcus admires.

I can edit my first message so that it instead has

Kaj: Here's a scene with a teenager with his parents. 

[scene with Marcus explaining differential equations, edited]

Write a follow-up to this scene where Marcus is anxious about whether a girl at school will like him or not. He thinks that she's both pretty and kind, and also into environmental activism, which Marcus admires.

This also makes the conversation shorter, which helps in not running up against usage limits so quickly. Since the initial scenes do a lot to establish the mood, I might spend some substantial amount of time on them - they might often be a mixture of my own writing and specifically-prompted LLM writing, with excerpts from several different LLM-produced outputs stitched together.

Making additions to existing content

One way to use an LLM is to make revisions to old content, either to incorporate new material that I wouldn’t be sure how to write or to just spice something up.

In one of my stories, it became established partway into it that a couple of the protagonists had received specialized training in reading things like microexpressions and body language and were very good at it. However, I hadn’t known this about the characters when I started writing, so I had several scenes that didn’t show this in any way. That felt off, so I wanted to go back to those earlier scenes and revise them.

I decided to ask Claude to do an initial editing pass on those scenes to incorporate this. Here is an excerpt from the first scene of the original story:

Imperial Baseship Terion, flagship of the 86th Expeditionary Fleet. Qaix system.

"Okay, dear", Admiral Rayne said softly, turning to look her catgirl in the eyes. "I think we can start returning to our quarters now."

Faye nodded with a smile, munching on a sugar donut as she did. They were at a formal ball for the planetary governor whose world they would be leaving behind tomorrow. Faye had been her usual charming self, clinging close to Rayne and looking around her with curious eyes and with her large cat ears turning about. She'd generally been shy but happy, letting Rayne do the talking and giving everyone friendly looks, generally making everyone around her cheer up a little.

Here’s a paragraph from a version that Claude gave me when I asked it to rewrite so as to show that Faye is exceptionally good at reading microexpressions, body language, etc.:

"Okay, dear," Admiral Rayne said softly, turning to look her catgirl in the eyes. She watched Faye's pupils dilate slightly—reading something in Rayne's face that prompted an immediate nod and smile.

Now, there are immediately issues with this. Claude is supposed to show Faye as having a good ability to read people’s expressions, but throughout the rewrite, it keeps slipping into having her owner (Admiral Rayne) be good at reading them (the bit about a slight dilation in Faye’s pupils). 

This feels like a very typical LLM kind of thing to do - you tell it “make character X better at reading people”, and it seems to hear something like “increase the general reading-people-ness vibe of the scene where the character is”. As a result, everyone in the scene gets better at reading people.

I’m also not quite sure of how the physicality works out - if someone has an “immediate nod”, there’s going to be a movement of their eyes away from you. Would you have the chance to catch a slight dilation of their pupils in that case? Also looking in a different direction will change the amount of light getting into their eyes and that would also affect pupil size…

But there’s still something good about this - notice how Claude has dropped Rayne’s bit of dialogue about them being ready to leave now, implying that it doesn’t need to be said as Faye can read Rayne’s intent even without it. That is something that I think works to communicate Faye’s skills pretty well, though it might also be a little too subtle - I’m not sure if the typical reader (or Claude itself, if I later feed it this scene in isolation) would catch that intent without it being pointed out.

So overall, I think that Claude is going in the right direction with the bit about Faye just immediately getting what Rayne is about to say, but the bit about pupil dilation has to go. Here’s how I ended up rewriting Claude’s rewrite:

"Okay, dear," Admiral Rayne said softly, turning to look her catgirl in the eyes. That was enough to prompt an immediate nod and smile from Faye, as well as a happy swish of her tail.

The extent to which the catgirl could read her owner felt almost supernatural at times. There’s this particular look you always get when you’re about to tell me we’re leaving an official event, Faye had once explained it. It’s obvious you’re about to tell me something nice, that’s not just about giving me more sweets.

You might notice that my rewrite of Claude’s rewrite makes it pretty different. But even if I’m discarding most of the details it added, the central idea that it gave me - that Faye would be able to tell Rayne’s intent from just a couple of words and some body language - was valuable.

Here’s Claude’s rewritten version of the next paragraph:

Faye munched on her sugar donut as they stood at the formal ball for the planetary governor whose world they would be leaving behind tomorrow. Faye had been her usual charming self, clinging close to Rayne and looking around with curious eyes, her large cat ears turning about. She'd generally been shy but happy, letting Rayne do the talking while giving everyone friendly looks—though Rayne noticed how Faye's tail would twitch whenever someone's smile didn't reach their eyes, or how she'd press closer when detecting hostility that others masked with politeness.

It’s a good start, but it is again slipping into implying that Rayne herself has a similar level of ability to read microexpressions as Faye does - how would she know what the exact signs are that Faye is reacting to?

That said, noticing that someone’s smile doesn’t reach their eyes is something that I would expect a person in Rayne’s rank and position to be able to do, especially if it’s pointed out to them. So I’ll edit the paragraph to suggest that Faye is giving Rayne subtle hints of what to pay attention to, and that Rayne is also perceptive, even if not quite as perceptive as Faye:

She'd generally been shy but happy, letting Rayne do the talking while giving everyone friendly looks. Though sometimes Faye would squeeze Rayne’s hand, such as when the admiral was speaking to the planet’s finance minister, and then Rayne would notice that the minister’s previously warm smile had suddenly stopped reaching his eyes. And on another occasion, Faye had drawn just a bit closer to her owner, and then Rayne had known that the defense minister approaching them was hiding hostility under his politeness.

This also allowed me to suggest that Rayne and Faye been working together for a while without needing to explicitly state it, which is nice.

Later in the same scene, the original goes:

"You did so well tonight, little catgirl", Rayne said as they reached their quarters, the tension in her shoulders easing as the doors sealed behind them. "I saw the governor’s aide trying to pet your ears and you handled it perfectly. Now, what did you hear?"

Faye nodded eagerly. "Lots of things. I think what you want to know right away is that the same aide was whispering with the governor, it sounded like they were happy to have us leave. 'Just in time', they said, like they'd been worried about us sticking around for longer."

"Great, that's important information. I'll have a stealth squadron left behind, see if they would pick up on what it is that the governor doesn't want us to see. Anything else urgent?"

In rewriting the scene, Claude takes the initiative to add some additional dialogue as well:

"You did so well tonight, little catgirl," Rayne said as they reached their quarters, the tension in her shoulders easing as the doors sealed behind them. "I saw the governor’s aide trying to pet your ears. You handled it perfectly—though I'm curious what his hand tremor told you about his real intentions."

Faye nodded eagerly, her pupils dilating slightly—the way they always did when she was about to share something significant. "His fingers were twitching in a specific pattern, like he was nervous about more than just touching me. Also, what you want to know right away is that the same aide was whispering with the governor. It sounded like they were happy to have us leave. 'Just in time,' they said, like they'd been worried about us sticking around for longer. But the governor's breathing changed when he said it—shallow, quick. He's scared of something specific."

"Excellent. That's important information." Rayne made a mental note of how Faye had prioritized the intelligence—sound first, then the body language that gave it context. "I'll have a stealth squadron left behind, see if they can pick up on what has the governor sweating through his formal jacket. Anything else urgent?"

Ugh. This dialogue feels unnatural to me - it has an info dump kind of quality, with the characters saying things for the benefit of the reader. I don’t think that an admiral asking for intelligence would specify that she’s interested in the hand tremor in particular, especially since Faye could have been reading some other signs at the same time. Rayne’s mental note also doesn’t make any sense to me, and I don’t like the way Claude randomly changed my “what the governor doesn’t want us to see” to “what has the governor sweating through his formal jacket”.

Still, the idea that Faye derived some non-obvious information from the exact way that the aide touched her makes sense. So does the bit that she intuited something from the exact way in which the governor said ‘just in time’. 

I’m however not entirely sure if the “he’s scared of something specific” bit works - doesn’t “Just in time,' they said, like they'd been worried about us sticking around for longer” already imply that the governor was afraid of a some specific thing happening if the fleet stuck around a bit longer? 

Then again, I could easily imagine a situation where a local governor gets generally nervous about “the Space Feds” showing up and hanging around. Maybe the Space Feds have a habit of throwing their weight around and making everything worse for the locals, regardless of whether the locals have done anything to justify it.

I’ll try to edit this to make it more apparent. This also gives me an opportunity to clarify the bit about “the stealth squadron” in the original, that was mostly there as a throwaway piece of flavor that I hadn’t put very much thought into. My rewritten rewrite reads as follows:

"You did so well tonight, little catgirl", Rayne said as they reached their quarters, the tension in her shoulders easing as the doors sealed behind them. "I saw the governor’s aide trying to pet your ears, and you handled it perfectly. Now, what did you observe?"

Faye nodded eagerly. "Lots of things. The aide’s fingers were twitching when he touched me - he was nervous about something, and trying to soothe himself by petting me. Also, what you want to know right away is that the same aide was whispering with the governor. It sounded like they were happy to have us leave. And there was something about the governor… as if he was scared of something specific happening if we don’t leave soon. 'Just in time', he said."

"Excellent. That's important information." It wasn’t unusual for planetary governors to get nervous when an expeditionary fleet showed up. But nervous was different from scared, and Faye was precise about getting those kinds of distinctions right. There was something that the governor really didn’t want them to know about… something that might be detectable from orbit. 

"I'll have a stealth squadron left behind. See if they can pick up what the governor doesn’t want us to see. Anything else urgent?"

While this has still required some editing and rewriting, it has still taken much less effort than trying to come up with all the tells of Faye’s perceptiveness entirely from scratch.

Spicing up writing

Another similar use case is to spice up writing that’s otherwise a bit monotonous. For the same story, I had written a flashback scene that showed an introductory speech Faye’s cohort of catkin got once they were fresh off their nursery and into their training program, before they were ready to serve at the sides of the officers. I thought it was okay content-wise, but it was a bit dull - just one character talking in a wall of text fashion.

This was what I’d written originally, without AI involvement:

(Content warning for: institutional indoctrination of adolescent catkin, allusions to physical punishment including electric shocks, cruelty hidden in a pleasant tone.)

I asked Claude to rewrite it in a way that would spice it up while keeping the original content, and it gave me an edited version. Now, the first version that it gave me was laying it on a bit too thick, and I deleted a number of its additions. Here’s a version where I’ve deleted some new stuff that I didn’t like, but haven’t otherwise edited it:

I might still edit this a bit to change some of the wordings, but overall I like this better than my original. It now gives a much stronger sense of the overall vibe of the room, and the way that their moods are shifting as Patel is talking.

Establishing a co-writing persona

By default, Claude’s voice when brainstorming for story ideas feels a bit detached and clinical. It sounds like, well, an AI. It might call an idea ‘compelling’ or ‘intricate’ or whatever, but it doesn’t really sound like it’s personally excited about it. It’s just there to help me come up with stories, just like it’s there to help me with everything else.

To change this, I often give it an explicit writer persona. The original motivation for this was that I might often start stories but then notice that I would eventually run out of ideas for them and then get bored with the story, starting a new one. I had noticed that with human co-writers, this was less likely to happen, as humans would have their own ideas and bring in fresh perspectives to complement mine. So I experimented with a few prompts that were aimed at getting it to suggest more ideas.

The persona that I currently use the most is a slight variant of this one (technically I have this as a “Style” within Claude, so I can enable and disable it at will).

I’m not sure exactly how much of a difference this persona makes to the actual prose itself, or whether it makes the writing any better. In fact, sometimes when I try generating messages with this style either enabled or disabled, I find that I prefer the ones where it wasn't on.

However, I find that this one is a lot more fun to write with than the default style. For one, it perfectly captures the slightly deranged vibe of many human writers (he said affectionately and self-deprecatingly) where we get actively excited about doing terrible things to our characters and go “OH MY GOD THAT WAS HORRIBLE” about something we just wrote… and then go on to do more of it.

It’s actually kinda hilarious how excited Claude with this persona acts about writing dark stuff.

Now, I do suspect that if you are trying to do co-writing with an LLM, it’s useful to explicitly try to get it into a persona that you feel matches your creative vibe. Co-writing is very much about trying to have matching vibes, and LLMs are already vibe-matching machines, so you might as well lean into it and make use of it. Human writers excitedly gushing to each other about the parts of the other person’s writing that they like is part of the mechanism of how they get their intentions synced up, and I’d expect the same to work with LLMs.

Often when I alternate writing messages with Claude, I put in little out-of-character observations and comments in my various messages. As a result, Claude will do the same, throwing in commentary and ideas on whatever we’re writing on the fly. I find that this tends to spark more ideas in me than if we had a strict division of “planning, then scene, then analysis and planning, then next scene” like I originally did.

The persona has some specific instructions about only writing its assigned characters and leaving my characters to me. I find that LLMs are not always good at following this instruction and typically have to be reminded at least once, or they’ll slip into writing all the characters (and might do it anyway).

At first I found this slightly annoying, because I was trying to do the same thing that I was used to doing with human writing partners, with a strict separation of characters. However, I soon found that once a story went on and both me and Claude got to know the characters better, I actually became more open to the characters being less strictly divided between us.

At first, it was just me going “I’d be curious to see you writing my character” when I was blocked for ideas and then using Claude’s take for inspiration, copying an edited version of it into my message. Then it gradually transformed into more fluid transitions, where I might mention “maybe the next scene could be something like…” in one of my messages and Claude would write a draft of it (writing both of our characters) and then we’d end up extensively iterating on it. 

Now I might still want control of specific characters in specific scenes if I have a particular idea for them, but we can also do a looser thing where we just alternate writing turns in general.

I've also experimented with some other custom styles. Another attempt at solving the problem of new ideas predominantly coming from me was a style that explicitly instructed Claude to consider the amount of open loops and to open new ones if there weren't enough. 

This worked somewhat but often did things like mention ominous footsteps coming from the corridor when there was no real reason for the story to suddenly take an ominous turn. Then again, if I want the AI to introduce new ideas, maybe I shouldn’t complain about it then taking the story in unexpected directions! So it might have worked with more tweaking, but so far I've ended up preferring the version where we just do plenty of out-of-character chatter and brainstorming as we go about it. 

Someone might else prefer the loop prompt, though - worth experimenting for what works with your writing style.

Thanks to Justis Mills for comments on a previous draft of this article.

This article was first published as a paid piece on my Substack one week ago. Most of my content becomes free eventually, but if you'd like me to write more often and to see my writing earlier, consider getting a subscription! If I get enough subscribers, I may be able to write much more regularly than I've done before.