Curate Your Own Pipeline
Issue #80: How to use LLMs and write like a centaur
Stanley Chen rewrites every line that LLMs generate for him. He believes current models are neither sufficient as writers nor as judges of literary merit. And yet, long before the public had heard of ChatGPT, he made LLMs an essential part of his fiction writing process. He’s enough of a power user to have opinions about which model best handles Chinese punctuation marks (Gemini), and whether it’s worth running models locally for novel-length text (no).
Few science fiction authors have been as involved at the frontier of technological innovation as Stanley has. In 2019 he co-wrote a book, AI 2041, with the CEO of Google China, Kai-Fu Lee. That same year, the World Economic Forum tapped him to help imagine possible futures. I had the opportunity to ask Stanley about his process and his work. What follows are some highlights and you can find the full transcript here. The text is lightly edited for clarity. All quotes are marked with timestamps [min:sec].
Protocolized encourages its writers to use LLMs and other AI tools. I’ve been trying to use LLMs in my writing, but just haven’t been satisfied with the results recently. At the same time, I feel an ongoing tension about that – how could I not be finding it useful to draw upon the entire combined archives that were scraped from the open web?
Stanley touched on this feeling early in our conversation:
[10:20] I think it’s not there yet. So you feel like always there’s something missing here, or this character might not seem to be 100% aligned and this dialogue could be a little bit cheesy. There’s always something there – so you will not be satisfied at all.
Despite that, he’s designed a custom workflow that he feels does enhance both his fiction writing and his fiction translation outputs. As a longtime teacher of fiction writing, he’s also able to provide some direction for what to focus on to create actual improvement. Indeed, our conversation did frequently return to the taste that needs to underlie successful LLM use, and his suggestions for how to develop it. Be deliberate, Stanley says, about how they are prompting you:
[03:56] It’s not just you prompting the LLM, but also the LLM as a prompt reversibly triggers your imagination and creativity.
Where LLMs fit into Stanley’s writing process
Given the amount of time Stanley has spent tuning his use of LLMs for writing, I asked:
[3:02] Where within your writing process do you see there being a use case for LLMs? Do you see it as more relevant during the writing process or editing process? What’s your philosophy there?
His response:
I thought it was just for collecting the data and information and trying to build the structure. But now I think it’s it’s both – it’s when you’re conjuring ideas, when you use it for world building, character building, and structure building. But also you can use it for rendering your style in a way and trying to find a correspondence between different disciplines. For example, when you try to use a very comprehensive metaphor to describe some super complex science theory. An LLM can usually give you some options that you can choose from…
I will use different models for different purposes, and I will separate different tasks. Maybe one big task, you can break it down to different small parts and you give it to this model or give it to that model, or you might want to rerun an output in a different model1 And you try to find the best outcome. But you have to, I mean, you have to train yourself with some kind of good taste.
This question of taste reappeared throughout the process, craft, and philosophy sections of our conversation. Taste and agency have been two buzzwords of 2025, for anyone tech-adjacent and terminally online.
My question to Stanley, then, was: Can you improve your craft as a fiction writer using LLMs? If so, how? This was a question that was particularly on my mind as someone relatively new to fiction writing. It seems like a skill issue not to have found a good workflow for this. Stanley agreed – improving your craft happens beyond the actual writing practice. An LLM can give you options, and your job as the writer is to have a consistent taste to discern which among them serves your vision.
That means no copy-pasting.
[10:04] I always rewrite. I couldn’t use even a single line it gave me directly. I think in a way there’s some legal issues. But through time and through your career, you raise the bar. You have a higher understanding, you have a deeper understanding of writing as a craft. You have to redo it over and over again. When you find your touch, you find the right feeling. I think that’s the most important thing.
This really reminded me of a college professor I had who swore by manually retyping code snippets from Stack Overflow. “You’ll see what’s wrong with it and change it when you rewrite it,” he said.
My former Spanish teacher said something similar. His rule was that when using Google Translate, even if you know the translation it gives you is correct, you still have to re-type it yourself. And then when you do, you often notice ways in which it was subtly wrong.
Oh, and speaking of translation –
Translating fiction with AI
Stanley and Ken Liu have had a multi-year, two-way partnership in translation – Ken Liu translated Stanley’s Waste Tide into English. Stanley translated some of Ken Liu’s work into Chinese.
[11:46] When I translated some of Ken Liu’s work into Chinese I used an LLM to find the tongue. I think it’s very difficult to capture the style, the tongue. It’s at the inter-linguistic level. You have to find the right slang, for example, or the right metaphor. Sometimes there are some cultural references that might not be able to directly attach with something in Chinese, or the other way around. So I think it’s always useful to use LLMs to help you to find something you know as a counterpart of the the language and the culture. Again, this is about taste. If you are bilingual you can feel the nuance and you can feel the subtlety of the language and the literature and you you might feel, “oh this is not exactly what the author tried to do in the original version.”
So sometimes they do experiments as well. The previous work can be referenced, but it’s not exactly the same, right? So we have to... dive deeper and we try to understand what the author really tried to do. And we come up with this idea, or maybe we could create this kind of style in Chinese to fully capture what the author tried to do. And we try to have the LLM represent what we think might be correct or close enough. But again, it’s not easy. So maybeit’ll give you the mediocre, or an average level of translation.
So I think this is like back and forth. I don’t think you can fully rely on LLM to do the translation, especially on literature.
Again, the role of the LLM here, in Stanley’s view, is largely to present you with a palette of options.
So what’s happening in between, then? What’s happening while you’re redoing it over and over again, prompting the LLMs and letting them prompt you back?
[28:48] As with every tool, you can use it in a dumb way, or in a smart way. You can use it to be lazy, or you can use it to challenge yourself. You think firstly and you see what the AI gives you and you try to compare the two. And you can see, okay, maybe this is what I’m missing here. These are my blind spots and there’s something nice, nicely done there. So maybe this is the focus area I should look into next time. So I think it’s a very good practice for your brain, for your imagination, for your narrative...
This was the thread I was really trying to pull on throughout our entire conversation. What does it look like, to Stanley, to use LLMs in a smart way?
It means being in dialogue with them, and being fluid about their role in your process. I asked Stanley what his mental model is when he uses AI in his fiction process – tool, palette, partner? Something else?
[4:52] It’s inter-subjectivity, so it’s something in between. You can see this subjectivity shifting or flowing between me and the machine in a distributed way. So I think all the magic happens in between. During the process, you know, think about electricity, think about the energy flows.
Interacting with your characters
Generative technologies offer new opportunities to interact with your characters and to discover them, even as the author.
[15:05] I’ve been playing with Midjourney and, recently, with Sora 2. It’s very interesting because when you have a character in your imagination there’s a distance, right. It’s kind of vague, kind of abstract. But when you transform it into a very vivid and and visually stunning image, even an animation, you give it more information. You might be triggered with a sense of how the character acts, or talks, or responds to certain scenarios.
What does it look like, though, to flesh out those characters, and to visualize them within the space of Sora or similar tools? Stanley experimented with assigning characters into the 16 Meyers-Briggs personality types.
[16:49] Now you can divide yourself. You can role play with different personalities. Each type could be attached to specific characters. You can interact and this becomes a role-playing game, like LARP. But also I think in the future, we might use Midjourney or Sora to do a real role-playing game. Everyone could create something like that. It’s not only writing a novel, but also creating something visually. I think that’s the future.
He’s been involved in an immersive drama production in Hangzhou. The premise is that its eight characters are all reincarnated.
[19:25] We have previous lives, we have this life, and the next life. When people buy the ticket, if they pay more they can see different lives. If they pay less they only see this life. So they can choose, just like in Sleep No More in New York. And you can follow different story lines to explore. I found it fascinating because there are some interactions between the characters and the audience. And it might trigger something that wasn’t in the script. Sometimes the character might have to improvise. But from the very beginning, we had like a big Bible defining the world building, and each character, like attached to the different lives. Part of their identity and part of their personality has been changed, like your job, your career, even your appearance, but part of it, like the soul, you know, like some addiction, like some fears, like some anxiety pattern has not. This is a very different way of creating and also a different way of storytelling.
Keeping the process separate from the product
However, Stanley cautions, you still need to be deliberate about how much information you are providing to your readers.
Unveiling the process is helpful to a point, but the work needs to stand alone. The artist should clearly have a vision about where the work ends and what the reader needs to figure out on their own.
This is one of the aspects I admire most about Stanley’s fiction. As much as the process is important, and as much as he’s been willing to publicly share about it – including in this interview – the work itself is the thing.
[21:07] Maybe there’s some fan fiction – readers can develop the character by themselves, they might have some alternative history, some subplots – like branches, storylines. So they can develop something like that. But I think to me what I should reveal to the readers is already written. You can interpret it by yourself, that’s your freedom. I always say, OK, just figure it out by yourself because what I want to say is already written there.
Even the endings. With a semi-open ending a lot of people ask me, “what does it mean?”
You have to fill in the blank with your own imagination. I think that’s the best part of it. You can keep guessing, you can keep thinking about it. That’s what I try to do.
Shaping the future with writing
Part of shaping the future with writing is about experimenting with form and embracing the new.
I asked about the prospects for young, emerging authors:
[48:59] They are suffering because there’s no platform at all. There’s no publication. There’s no opportunity for them. So I think this is very, very cruel. People feel angry. People feel that they’re outliers, that they’ve become irrelevant in a way, right. I don’t see an easy solution here.
But I always told my students, you have to create your own job in the market. Maybe it doesn’t exist yet, but you have to start to get down to work and start to build your own career, not just waiting there, waiting for some companies, you know, some bosses to give you one, to offer you one. I think this is the reality.
And with AI, I think, yes, they can do it, but they need to refresh their experience. They have to unlearn whatever they’ve been taught, you know, from the traditional schools or parents or educational institutions.
In his own work, Stanley has been very proactive about forming partnerships with people directly making decisions that impact the future.
While he’s a sci-fi writer, he has been invited give guidance in technology and economics discourses.
So what perspectives does he find it useful to share in those forums?
[56:36} We’re so small, we’re so powerless. This is a structural asymmetry. I think punk, like solar punk, basically is what we need here. And how we should do that starts from creating new kinds of metaphors. Because metaphor is how we package and compress all this storytelling, emotion, ethics, values, mythology all together. It’s like a super efficient info pack. But it takes a very long time to form a metaphor. So that’s why we should start now. We need some time to build up new metaphors…
I try to shift the vision for these people because they are the one who spend the money, make the action and policy-making, etc. I think you should do something to change the narrative, change the metaphor and tell a new story and create a new myth for the future.
Lastly, I couldn’t resist asking about some specific procedural steps:
[7:40] I think Claude is good on the language level – it is more nuanced, is more human in a sense. But I found it couldn’t do full-width punctuation marks in Chinese. I’ve tried different ways to teach it how to use the right marks, but I fail all the time. Gemini has a longer context window so it allows us to have novel length materials. And you try to summarize something or you try to dig deeper in a very accurate place, even down to sentence level. Gemini helps a lot in doing this kind of long text task. So I think, of course, different models, they have different characters and they have their different styles.
You have to really dig deeper and try to find your own pipeline. I think this is very unique individually. Design and curate your own pipeline to better optimize your workflow. This is how I adapt myself to this new way of working, new way of writing.
Listen to the full conversation that I had with Stanley:
 | A guest post by
|