For a while now, I have been really enjoying co-writing fiction together with LLMs.
There is a preconception that LLM-written prose is bad. And if you only give the LLM a two-sentence prompt and then have it do all the work itself, it often is! But if you use a good model, throw your soul into it by telling it exactly what you want, craft an initial scene to build upon, and aggressively rewrite its outputs to address issues, it can become much better. (That said, this post will not try to argue for that or show it and will just take it as a given; see Justis's earlier post for some examples if you're curious.)
For a decade or two, I've been having difficulty finding things that I could do on my own that would consistently bring me pleasure. This has been a major contributor to varying levels of depressive episodes. Even in the absence of anything clearly bad going on in your life, just the mere lack of anything consistently good that you could feel unambiguously excited about can bring with it a sense of pointlessness.
With LLM co-writing, I've finally found such a source of pleasure again. I've occasionally been finding myself waking up in the morning and feeling actually excited about getting to continue some story that I'd been writing before. It has felt like something has been coming alive - some deep emotional content finding a way to be expressed in a deeply satisfying way.
However, it's taken some time to get to a point where I really enjoy it, and I think I'm still in the process of figuring out the best ways for it. In this article and my next one, I'll try to share some of what I've learned.
Why write with an AI?
If you are happy writing your fiction yourself and have no interest in co-writing with an LLM, then you indeed have no reason to do so. You do you.
It used to be that I really liked writing fiction, but couldn't write longer stories myself. I'd manage a couple of scenes and then just run out of ideas, or feel like my characters didn't come properly alive. My style of writing tends to involve immersing myself in the head of one character and writing from their perspective. This has the problem that it makes it hard to jump into the heads of the other characters at the same time, so they end up flat and lifeless.
This problem can be avoided if I have a co-writer. They write one character, I write another, and then we have those characters interact. I also love the interplay of ideas when writing with somebody - I introduce an element into the story and see how the other person builds on it, and then they introduce an element of their own and I build on it. It’s amazingly fun to come up with something and see what someone else does with it.
When I was younger, I wrote together with various people, and at best we might do intense writing marathons, emerging half a day and tens of pages later. But it can be difficult to find a good writing partner, especially once everyone tends to have adult responsibilities.
For me, co-writing works off momentum. It takes some investment to drop into that character headspace, and once I’m there, I don’t want to leave. I’ve had several brief attempts at co-writing together with someone that ended because one of us couldn’t commit enough time and energy to make it work. Or worse, it did work and I got obsessed with it, but then my partner couldn’t invest a corresponding amount of time into it and the collaboration sputtered away just as I’d gotten excited about it.
It can also be challenging to find someone whose tastes and writing style fit together with yours well enough. It's of little benefit if your schedules would work together, but the things you'd be excited to write about just don't interest them, or if their prose leaves you feeling uninspired. And it might also be that you get really excited about some direction where the story is going and would want to pursue it, only to find out that they find it uninteresting or even off-putting.
I'd had enough of these kinds of experiences, as well as attempts to write together with someone that ultimately ended up going nowhere, that I basically stopped trying to find writing partners.
But if I write together with an LLM, I don’t need to try to figure out how to make our schedules fit together; I can just write when it suits me. The LLM is reliably available, and it doesn’t get bored if I want to iterate on countless slight variations of some particular theme that happens to be a personal fixation.
Is it as good as writing with a human? Well, it has many shortcomings that a human doesn’t have, but it also has strengths that a human doesn’t.
As an example, there is a common effect where people find it much easier to rewrite someone else's draft than to write something entirely from scratch themselves. I find that this is the case for me and LLM writing. Sometimes I ask the LLM to write a particular kind of scene and it's pretty good, and I'm content with making just a few edits. At other times, I get a scene and it's not at all what I wanted - but it gives me a clearer idea of what exactly I do want, enough that I can write an entirely new version of the scene myself! And sometimes I have the LLM produce three different versions of the same or very similar prompts, and then I use copy-paste to stitch together a Frankenstein's monster version of the scene that incorporates the best parts of all the versions... plus some of my own writing on top.
I guess it'd be possible in principle to do this kind of co-writing with another human, but it feels more like the kind of thing you'd do if you were working on something like a movie or TV script for money. If you're just writing with a friend for mutual fun, going "your writing sucks but give me three versions of that same scene and I'll combine them into one actually good one" would be a pretty dick move.
Meanwhile, a human can proactively introduce new directions and feel more intrinsically meaningful to write together than an LLM. So both human co-writing and LLM co-writing are great in their own ways, and I wouldn't say that one is better or worse than the other.
I have mostly been writing for my own pleasure. The things I've been writing have been quite idiosyncratic to my interests and preferences, and I don't expect many people to necessarily appreciate them. That said, I do have at least one story that I've been starting to feel has enough literary merit that I might want to share it more widely someday. I've been editing it as I write, but for now, that's out of a pure love of the craft - I just enjoy working on it for its own sake, regardless of whether I'll ever publish it.
In order to keep things simple, the stories used as examples here are all quite short. Most of the ones I've been writing are substantially longer - the one that I mentioned currently runs at about 20K words, and I expect it to still grow quite a bit from there.
The basics of getting good writing out of LLMs
Before we get to the bit about co-writing, let’s talk about prompting for fiction in general.
As far as I know, the best LLM to use for this kind of thing is Claude Opus 4. The Claude Sonnet models are also good. I’ve also heard people say good things about DeepSeek and GPT-4.5, though based on my brief experiments, I prefer Opus.
When prompting LLMs, a useful rule-of-thumb is to ask, "how well would a human writer do with this prompt?". For example:
"Write a short story about a relationship." - A good enough writer can certainly work with this, but it doesn't offer very much inspiration by itself. Many people would struggle to get started.
"Write a story that might get accepted at a prestigious literary magazine." - I think in the typical case, a person hearing this instruction will think of the kinds of stories they've read at prestigious literary magazines (assuming they’ve read any) and then write something that mimics the superficial style of those stories. Whatever they produce as a result is unlikely to be accepted at such magazines, because that’s not how you write a good story.
You write a good story by thinking about cool characters, themes and ideas that excite you, lovingly crafting prose that you find beautiful, et cetera. Of course, you may consider where your story might be published and that may shape the choices you make around it, but the core of the narrative has to be something deeper.
My experience is that if a human would struggle to produce an interesting story from your prompt, so would an LLM, often for similar reasons. On the other hand, if you, e.g., give it nuanced characters beforehand and describe some of their dynamics, you can get something much better. Most of my writing with Claude starts with several messages brainstorming or discussing character details together; then a scene with those characters; then a few messages analyzing the scene together, establishing new facts about the characters and the setting, then a new scene again, etc.
For example, here’s the beginning of one conversation with Claude Opus:
Kaj: I was thinking of writing some story together with you, some kind of psychological chess game perhaps
Hmm "psychological chess game" makes me think some high fantasy evil overlord. Like... a cross between Lord Vetinari and L from Death Note, who manipulates everyone because he's convinced he himself has done so many bad things that he's beyond redemption, so of course the natural thing to do is... do more bad stuff. Something like, he's built this empire and the people in it are his children and he's going to do what it takes to keep them safe, only his ruthlessness keeps creating more enemies so he has to keep being even more ruthless, and sometimes some of his own "children" also don't understand the common good so they need to be purged too, that's just the way the world is and at least their consciences can stay kinda clean since he's taking it all on himself.
Maybe he's got this protege, someone whose life he saved when she was little, and then he saw her potential and she became a bit of a daughter figure he's been preparing for serving as his right hand - only so far she's had this very idealized view of him and now it's time to start preparing her for some more ruthless work.
Claude Opus: [You have] the ruler who justifies increasing brutality through a twisted sense of martyrdom, and the protégé whose idealized view is about to shatter. The "I'm already damned so I'll damn myself further to protect others" logic creates [...] spiral where his protection becomes the very thing that endangers them [...]
... the protégé's awakening [is] the potential story catalyst. She's been living in a carefully curated reality, and now he needs her to see behind the curtain - but how much can he reveal without losing her loyalty entirely?
Are you imagining this as the opening conflict - her first "real" assignment that forces her to see what the empire actually runs on? Or were you thinking of building up to that moment of disillusionment more gradually?
Roughly this alone can be enough for Claude to write something pretty good. Here is a five-page scene that it created based on a version of the same conversation. However, I usually like to do a little more plotting than this, so let’s continue with the planning.
Claude’s question of whether this would be the opening conflict gets me to reflect that it might not be. I think about what the alternative would be, and hit upon the idea that if the protégé has fought hard to get where she is, that would create a very strong internal conflict if she were to ever question the overlord.
Claude takes the idea and runs with it. It ends up suggesting several interesting ideas - that we might show her in training and seeing other trainees failing and disappearing, how she might suddenly start noticing small inconsistencies but rationalize them away, and then get increasingly deeper into rationalizing away everything that might shake her faith in the overlord. It also asks me some more specific questions that give more ideas.
Kaj: I was originally thinking of this as the opening, but it might be more interesting if we build up to it gradually. Any ideas of the central conflicts happening before that - maybe she has felt driven to prove herself in his eyes earlier and is afraid that he won't see her as impressive enough? Like maybe it wasn't always obvious that he'd hold her in such high regard, she actually had to earn it. That would then make it hit even harder when she starts glimpsing at the truth - she'll have to question whether it was all worth it, and the same time the fact that she's invested so heavily in it means that she'll feel pressurized to rationalize everything he does as okay.
Claude Opus: [...] If she's spent years climbing toward his approval, every glimpse of truth becomes a threat to her entire identity - not just her view of him, but her sense of self-worth tied to earning his regard.
I'm seeing a few possible stages in her earlier development:
Maybe she starts as one of many talented prospects, competing for his attention through increasingly difficult trials. She excels at strategy and analysis, but initially he seems unimpressed - perhaps deliberately, to push her harder. She might notice how other proteges disappear after failing certain tests, assuming they were simply dismissed or reassigned.
There could be a pivotal moment where she does something unexpected - maybe she solves a problem through empathy rather than ruthlessness, and instead of punishing this "weakness," he suddenly shows interest. She misreads this as him valuing her compassion, when really he sees potential to mold someone who can inspire loyalty in others while he handles the darker necessities.
As she rises in his estimation, she might start noticing small inconsistencies - orders that don't quite match official explanations, advisors who seem fearful despite their privileges. But she's gotten so good at explaining away his actions ("He has access to information I don't," "These hard choices prevent worse outcomes") that the rationalization becomes automatic.
What kind of early "tests" were you imagining he might have put her through? Things that seemed like leadership challenges at the time but might look different in retrospect?
I can also approach the conversation from a different frame: giving the same initial description, but asking for a psychological analysis rather than approaching it from a storytelling perspective. I often like to do this since I enjoy starting from detailed and complex characters.
Let’s see what happens if I use the psychological frame rather than the narrative one. I give the same initial description about the overlord and his protégé, but instead of saying "I was thinking of writing some story together with you", I say "Please analyze the psychology of these two characters, drawing out any implications that haven't been explicitly stated, as well as suggesting possible other details that would fit together with what's already been established."
Claude Opus: The overlord's psychology operates on a [...] self-fulfilling prophecy - he's created a moral framework where his own perceived irredeemability becomes the justification for continued moral transgressions. This suggests someone who experienced a formative betrayal or failure early on, something that shattered his ability to see redemption as possible for himself. The comparison to Vetinari and L points to someone who views human nature as fundamentally predictable and manipulable, but unlike Vetinari's pragmatic stability-seeking, he's trapped in an escalating cycle.
His paternalistic view of his subjects as "children" reveals a deep need for control rooted in the belief that others are incapable of making correct choices without his intervention. This worldview likely emerged from witnessing chaos or suffering that he believes could have been prevented through sufficient control and foresight. The fact that he must periodically "purge" his own people suggests he's increasingly unable to distinguish between external threats and internal dissent - a paranoia that grows as his methods create the very resistance he seeks to prevent.
The protégé [...] simultaneously his attempt at legacy and his greatest vulnerability. By keeping her idealistic view intact until now, he's been protecting something pure in his world, perhaps the last remnant of who he might have been. The decision to begin her "real" education suggests either external pressures forcing his hand or an internal recognition that his mortality demands a successor who can maintain his system.
Her psychology would likely include an intense loyalty born from that childhood rescue, creating a savior complex that makes questioning him feel like betrayal. She probably possesses exceptional observational skills (explaining why he chose her) but has unconsciously trained herself not to see certain patterns in his behavior. Her idealized view isn't mere naivety - it's an active psychological defense against recognizing troubling truths she's half-glimpsed.
Some details that would deepen this dynamic:
[...] The protégé might have developed her own network of contacts who trust her because of her genuine belief in the overlord's vision, making her upcoming disillusionment potentially devastating to the power structure. She could have a parallel relationship with someone she's mentoring, unconsciously recreating the dynamic she has with the overlord.
Claude is now suggesting lots of juicy backplot that we could explore. The overlord's "worldview likely emerged from witnessing chaos or suffering that he believes could have been prevented through sufficient control and foresight" - yes, that sounds exactly right. So exactly what chaos or suffering did he witness? Hmm.
I particularly like the bit that Claude suggested about the protégé having a group of people who trust her because of the strength of her vision in the overlord’s cause. I imagine that once she starts seeing the truth, that would create another internal conflict. On the other hand, she would not want to let her friends down by having her vision waver, but on the other hand it would feel like a violation of trust if she wasn’t honest to them about what she was learning. It’s a double bind where either way she acts, she feels like she’s betraying them.
When I mentioned this to Claude, it threw in more details, such as suggesting that protégé's network may consist of people who have also invested their sense of meaning into the overlord being right, so they depend on her being right. To me this seems like a particularly realistic dynamic to explore, given that I've seen many communities (such as LessWrong) seem to attract people whose sense of meaning resonates with that of the community's founders. So it makes sense that the protégé would also attract people who were drawn in by the strength of her vision and faith in the overlord.
But then that makes it so much harder to get out... and Claude suggests that the protégé might soon end up in a situation where the way she treats her followers starts to mirror some of the manipulation that the overlord is doing on her. Mmm. I like this. A very realistic-sounding description of how she might end up becoming exactly the same kind of person she's starting to rebel against.
Claude Opus: [...] Where he justifies increasingly dark actions through his perceived irredeemability, she's caught between two forms of betrayal - maintaining a false faith or shattering the faith of others. [...] both are trapped by the weight of others' dependence on them.
The protégé's network would likely consist of people who've also been helped or saved by the overlord's regime - perhaps orphans given purpose, former criminals offered redemption through service, or idealistic young administrators who see her as proof that the system rewards merit and loyalty. They don't just trust her; they need her to be right about the overlord because their entire sense of meaning depends on it.
As she begins seeing the truth, she might find herself unconsciously mimicking the overlord's own patterns - small lies to protect her followers' innocence, rationalizing that they're "not ready" for certain truths. Each small deception would make the next easier, even as it makes her feel complicit. She might even catch herself thinking "I'm already lying to them, so what's one more omission?" - echoing his logic about being beyond redemption.
This bit of conversation is also a good demonstration of the way that you get out reflects what you put in. I initially suggested a particular kind of psychological dynamic - that the protagonist will have invested her identity in winning over the overlord’s favor, and that she’ll feel pressurized to rationalize away any threats to that identity. Once I brought it up, Claude understood it perfectly and built upon it, suggesting a lot of specific details that fit together with that pattern. If we start writing now, it will keep that context in mind and incorporate details that fit together with that.
But none of that would have happened if I hadn’t suggested it. One way to think about it is that an LLM can write almost any kind of fiction, but it doesn’t know what kind of fiction you want it to write unless you tell it. If you don’t tell it anything, it will default to the kind of prose that the average reader might like - which can often mean relatively simple characters and trope-heavy stories, because those are what work for a mass audience.
Also, the LLM is constantly making background inferences of you in order to figure out what you’d like. Suppose that you tell it that you want one of the characters to be good at math, but don’t tell it anything more than that. Since you haven’t given it information that would indicate that you personally know math, a reasonable inference is that you don’t, so putting in real math details would just go over your head and possibly bore you. As a result, you might get a character that looks like a popular stereotype of a math genius, while only having a superficial resemblance to the kinds of people who actually know math.
Let’s try what happens if we do that:
Kaj: Write a scene involving a teenager who is a math genius, having breakfast with his parents
Claude Opus: The kitchen smelled like burnt toast and coffee. Marcus sat at the table, absently rotating his orange juice glass in perfect quarter-turns while staring at the condensation patterns forming on its surface.
"Did you sleep at all?" His mother set a plate of eggs in front of him, the fork clattering against the ceramic.
"Some." Marcus picked up the fork, calculating the optimal angle to cut through the eggs while minimizing the scraping sound against the plate. Twenty-three degrees. "I was working on something."
His father lowered the newspaper. "The Collatz conjecture again?"
"No. Something else." Marcus took a bite, chewed exactly sixteen times. "About prime gaps."
"You know Mrs. Chen called yesterday," his mother said, refilling her coffee mug. "She mentioned you haven't turned in your English essay."
Marcus's hand paused mid-cut. "I'll do it."
"When?"
"Today." He resumed eating, each movement precise. The eggs were getting cold. Temperature decreased at approximately 1.2 degrees Celsius per minute in this room.
The scene goes on, but I think we can already conclude that this is terrible. “Calculating the optimal angle to cut through the eggs”. “Temperature decreased at approximately 1.2 degrees Celsius per minute in this room”. A few references to famous open problems to show Marcus thinks himself capable of potentially solving them. This is just throwing in random math in a way that sounds sensible to an average person, while leaning heavily on a stereotype.
But let’s try this in a way where we give Claude something more specific to work with:
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 Opus: The kitchen smelled of coffee and slightly burnt toast. Marcus sat at the breakfast table, his calculus textbook propped against the orange juice pitcher, while his mother scrolled through her phone and his father read the sports section.
"Dad," Marcus said, pushing his scrambled eggs around his plate. "You know how you were asking about what we're doing in AP Calc?"
His father looked up, folding the newspaper. "The one where you complained about the homework for twenty minutes?"
"That was different. We're on differential equations now." Marcus grabbed a napkin and started sketching. "Okay, so remember how regular equations are like... snapshots? Like, y equals 2x plus 3 tells you exactly where y is for any x?"
"Sure." His father leaned forward slightly.
"Well, differential equations are about change. They're like—" Marcus paused, tapping the pen against the table. "Imagine you're tracking how fast your coffee cools down. The rate it cools depends on how hot it is right now, compared to the room temperature."
His mother glanced over. "Is this going to end with you explaining why I should reheat my coffee?"
"Mom, no. Listen—" Marcus drew a curve on the napkin. "So if your coffee is super hot, it cools fast at first. But as it gets closer to room temperature, it cools slower. The differential equation describes that relationship between the temperature and how fast it's changing."
His father picked up his own mug, considering it. "So instead of telling me the temperature at any time, it tells me..."
"It tells you the rule for how temperature changes. Like dy/dx equals negative k times the difference between y and room temp. Where y is the coffee temperature, and k is just some constant that depends on, I don't know, how good your mug insulates."
Again, it goes on, but this should be enough to give you a general picture.
Now, as Justis pointed out on an earlier draft of this article, this scene still has some stereotyped bits that don't really make sense. Marcus is at home, why would he start writing on a napkin with a pen instead of getting a paper and a pencil?
But the characterization is much better, even if the notation Marcus uses for the law of cooling is a bit unclear (but maybe that's just because he's explaining it on the fly). I can just manually edit this afterward and fix the silly bits. If you start with this kind of a scene, the LLM will have a much better understanding of how you want these characters depicted.
It also helps to ask for things that bring out some entirely different aspect of the character, to let the LLM know that you want a three-dimensional character, not one that only thinks about math and nothing else.
When I had this kind of an opening scene in a story, in the next scene Claude would still throw in mentions of stereotyped "math person" behaviors. I then took the original scene and manually edited in a mention of the character also enjoying a fantasy novel and being eager to write fanfiction about it. When I fed Claude the edited scene and asked it to create follow-ups, it would still incorporate the character's math interest, but in a less strawmannish way.
Besides manual editing, we can also explicitly ask Claude to write a second scene that explores another side of our characters:
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.
Claude Opus: Marcus stared at his physics homework, but the equations kept morphing into Emma Chen's handwriting from the climate action petition she'd passed around in English class. He'd signed it immediately, maybe too immediately.
"You're going to burn a hole through that paper," his mother said, setting a glass of water on his desk.
"Just checking my work." He flipped to a fresh page, trying to look busy.
She lingered in the doorway. "Everything okay?"
"Yeah, fine." The words came out too fast. His mother had that look—the one that meant she was cataloging every micro-expression on his face.
"How was school?"
"Fine."
"Marcus."
He put his pencil down. "There's this girl in my English class. Emma. She started this whole campaign to get the school to install solar panels."
His mother leaned against the doorframe. "The petition you mentioned last week?"
"She gave this presentation to the school board yesterday. Just walked in there with all these calculations about long-term savings and carbon offsets." Marcus picked at the corner of his textbook. "She actually got them to agree to a feasibility study."
"Sounds like she knows what she's doing."
"She's just—" He stopped, reorganizing his pencils. "In class today, she was explaining biomagnification to someone, how toxins concentrate up the food chain. But she used this example about orcas and PCBs that was way clearer than how our textbook explains it."
And so on. The LLM tends to draw on what’s been established and avoid deviating from the history and explicit prompts too much, unlike a human who might be bringing in all kinds of other details. So it helps that we’ve now introduced more details about Marcus. In the future, Claude is going to remember both that he likes math and the whole thing with Emma, and make use of both.
Most of the scene is pretty good, but there are places where it slips into not quite making sense. Later on there is this bit relating to the solar panels that Emma is trying to have installed at the school:
His phone buzzed. A message in the group chat from Emma: Meeting tomorrow at lunch to plan next steps. Need someone good with numbers to help model different panel configurations. Anyone interested?
Marcus stared at the message. Three dots appeared—someone else was typing. His stomach twisted. Before he could think too hard about it, he typed: I can help with the calculations.
Send.
It’s just one sentence in the middle of an extended scene, so it’s easy to miss if you’re not reading carefully. But if they are having solar panels installed on the school roof, would they really “need someone good with numbers to help model different panel configurations”? I don’t know much about installing solar panels, but I assume that there already exist standard configurations that you just use, and you’d probably hire a professional to install them for you anyway.
Unless… what if Emma has a crush on Marcus too, and came up with a slightly nonsensical excuse to get together with him! Sometimes you can relate to LLM incoherence with an attitude of “yes, and”, turning a lack of logic into creative inspiration:
Kaj: If they were having solar panels installed on the school roof, would they really be “modeling different panel configurations”? I don’t know much about installing solar panels, but I assume that there already exist standard configurations that you just use, and you’d probably hire a professional to install them for you anyway. Unless... does Emma have a crush on Marcus too, and now she came up with a slightly nonsensical excuse for them to get together? I bet she's busy trying to figure out what to say when Marcus does actually show up 😁
Claude Opus: You're absolutely right about the solar panel configurations! Most schools would just get quotes from professional installers who'd handle all that. I like your interpretation—Emma definitely seems to have created a rather convenient "technical problem" that requires Marcus's specific skills.
Here's what might be happening in Emma's room that evening:
—------------------------------------------------------------------------------
Emma stared at her laptop screen, which displayed three different solar installation company websites. Each one clearly stated: "Free consultation and custom configuration included."
"Okay, so maybe I don't actually need help modeling configurations," she muttered, closing the laptop.
Her best friend Jade, sprawled across Emma's bed, looked up from her phone. "You think?"
"It seemed believable at the time! He's always doing those complex calculations in class." Emma flopped back in her desk chair. "God, he's going to show up with graph paper and everything."
"Which is exactly what you wanted." Jade was grinning. "Miss 'I just happened to notice his organized calc work.'"
Emma threw a pencil at her. "I panicked. The group chat was just sitting there, and I saw he was online, and—"
"And you've been trying to talk to him for weeks."
"He signed my petition really fast." Emma spun in her chair. "Like, really fast. But then he just... walked away."
"So naturally, you invented a fake math emergency."
"It's not fake! We could... compare the quotes. Or something. Calculate the environmental impact?" Emma pulled up a spreadsheet. "I should at least have some numbers ready so it's not completely obvious."
I like this direction! It’s a good thing that Claude screwed up, because I wouldn’t have thought of this idea otherwise.
Still, there are often times when the LLM says something that just doesn’t make sense, or where it is so strongly influenced by stereotypes that it keeps slipping into them despite the instruction to do differently.
LLMs aren’t really capable of thinking from first principles in the way that humans do - rather they rely on cached patterns, that may sometimes be strong enough to override even things that you’ve established before. For example, it’s difficult for LLMs to write very young children, since there aren’t many realistic depictions of their internal psychology. Instead they tend to fall back into stereotyped descriptions of “children in general”, even if they’re explicitly asked to go for a realistic and age-appropriate description!
In the second part of this series, I'll go into more detail on what to do when an LLM is being dumb and consistently not doing what you want, as well as talking more about the actual co-writing aspects (rather than just brainstorming and prompting).
Thanks to Justis Mills for feedback on an earlier draft of this article.
1
1
1