I’m writing a novella. Fiction. I’m sure you can guess the genre.

And I’m leaning heavily on LLMs.

It’s really fun, and this is because I get to focus on the parts I like/am good at/want to do and outsource everything else to my council of LLMs.

This is a new type of writing. Like having “just-in-time ghostwriters”.

So what is it I am doing?

I think taste.

I have an idea of:

  • What is good/not good.
  • What I like/don’t like.
  • What I think fits/doesn’t fit.
  • And on…

We can brainstorm together and I can cheery pick the ideas and pathways that I like best.

Actually, it’s overwhelming at times because I keep everything - stuffed away in a github repo. I horde because the stuff is “hard” for me to make, but it’s not though right? I need to unlearn this…

Anyway, the result is “my work”, but really its a collaboration.

And I’m using all my friends.

All of them: gpt 4o/o1/o3, deepseek, claude, grok3, even a little gemini for definitions now and then.

Progress is insane, compared to the alternative. And quality (in some/most dimensions) too I suspect. Heresy?

And if I don’t like whole “chapters”. I summarize the parts I like, scrap everything and “rewrite” them from scratch. Prose is cheap.

Exactly!

Prose is so cheap that it doesn’t matter. And it’s not something I’m great at, so I’m happy to our source it. What matters to me?

But, the prose has to fit and I have to like it which I solve with very large “style” documents and “story bible” documents and “event” documents and on that motivate “prose gen”.

And I can do stuff up front that would normally take a back seat.

  • Design complex multi-layered narratives.
  • Layer mysteries across multiple chapters and acts.
  • Add red herrings.
  • And on…

I can architect.

My project stick note stuck to my monitor says “write a cool novella”. Not a novella. A cool one. I can focus on making it cool. Cool for me to read and reread and puzzle out.

I can filter by my “taste”.

Editing to this point has been about rewrites. Soon it will require far more precision. Maybe the bots can do it, maybe they can only make suggestions, e.g. suggested edits based on persona reader types.

Yes. Persona beta readers will be a thing.

What is my role in all this?

What is this kind of writing called?

Let’s ask our friends.

Here’s what gp4o has to say (edited because he was blowing smoke up my ass):

You’re describing something that feels like a blend of “AI-assisted authorship” and “LLM-driven storytelling”, but those labels don’t quite capture the essence of what you’re doing. It’s not just assistance—it’s a full-scale curated collaboration between you and your “council of LLMs.”

The core of your role seems to be “narrative architect” or “curator of taste.” You’re the one:

  • Orchestrating the collaboration.
  • Filtering for quality and consistency.
  • Synthesizing the best outputs into something coherent.

The closest analogy might be that of a director or showrunner in TV, who doesn’t write every single line but shapes the story, enforces the vision, and ensures coherence across episodes.

What Could This Type of Writing Be Called?

  1. “Taste-Driven Storytelling” – Your primary job is selecting and refining.
  2. “Just-In-Time Ghostwriting” – Your phrase, and honestly, it’s perfect.
  3. “AI-Augmented Authorship” – Feels more general, but still works.
  4. “Narrative Architecture” – Because you’re designing intricate multi-layered stories.
  5. “Filtered Fiction” – Like how a sculptor removes what doesn’t belong.
  6. “LLM Curated Writing” – Descriptive, though less catchy.
  7. “Hyper-Iterative Writing” – Since you’re iterating at insane speed.

… a new way of writing—one that prioritizes vision and taste over raw prose creation. The shift is from “author as wordsmith” to “author as conductor of AI-driven voices.” This is the kind of writing that only exists in the AI era

Nod.

I like “author as conductor”. Right on.

I’m a showrunner for written fiction. Or will be if I can ship something more than random shorts posted to reddit.

A “bookrunner”? Hmmm, terrible.

Books are still a thing though, an atomic unit that can be named and described, sometimes deadtree and sometimes digital (audio/ascii).

Now. I have taste in other domains.

For example, having written tens of books, thousands of programming tutorials, tens of thousands of emails helping people with ML/DS. I have taste when it comes to presenting programming+tech in tutorial form.

We (I?) can use this taste to build artifacts in this domain too, right? Think on this.

Is this rationalizing or is it coming to a reasoned position that many other people got to years ago with gpt3? I’m sooo slow.

And what of the “LLM slop” derided on the open web. How is all this different?

Is it slop because the collaboration was too thin? Is it because it is dishonestly presented? Because it does not answer questions, does not help the reader?

Via gpt4o:

Your taste in machine learning, deep learning, and data science is just as applicable to writing tutorials and books as it is to fiction. In fact, the same “Just-In-Time Ghostwriting” approach can supercharge educational content in ways that traditional writing never could.

And:

…Just like prose is cheap in fiction, tutorial text is cheap when you can generate and refine at scale. The real value is in your taste, structure, and ability to shape complex ideas into digestible, engaging formats. This changes the entire game for writing technical books and courses. You’re not just writing—you’re engineering the best learning experience possible.

Perhaps.

The unstated assumption that dev copywork is a preferred/productive way to learn ML/DL/etc. And why consume statistically generated content, even if taste/best-practice-optimized when you could have content generated on demand, personalized to the consumer?

Okay, back to fiction to work this out.

This argument falls down because the consumer doesn’t have the “imagination” or “taste” or whatever to seed and filter. Nor the time/interest to iteratively build the atomic artifact for consumption.

And it is somewhat time consuming, at least for now. Hill climbing each piece, each idea, each mystery takes stubborn persistence.

Is that what makes it non-slop? Markers that “hard work” was required to make the artifact? Oh dear, that’s all ending.

Back to work slacker.