Process

The writing journey can begin with any concept that sparks inspiration. The structured process, however, starts once a concrete idea is established for development.

1. Develop the story bible

   • Create a comprehensive guide to ensure consistency across all elements of the narrative.

2. Draft large language model prompts

   • Write the specific prompts designed to guide the large language model's output.

3. Generate scenes

   • Input the story bible, prompts, and ongoing scenes into the large language model to generate new content.

   • Rework and edit the generated scenes to verify they align with the story's overall tone and continuity.

4. Assemble the story

   • Collect all the generated and refined scenes to form the complete narrative.

The writing process generally operates as a concurrent pipeline rather than a strict sequence of steps. Although the story bible is largely completed before the others, the subsequent steps are executed incrementally. This involves a cycle of writing prompts for specific book sections, generating scenes from these prompts, and then immediately adding those generated scenes to the story. This contrasts with an approach where all prompts are written first, all scenes are generated next, and the entire story is assembled only at the end.

Inspiration might come from anywhere. It might be a cool idea I had while watching a show, which then rattles around my brain for years until I finally decide to do something with it. It might occur to me that two franchises share two elements which could serve as a single hook to form a story. It could be an idea I have while playing a game. I might have a roleplaying session with one or more characters with a generative AI, at the end of which I think might make for a good story.

Wherever the inspiration might come from, it remains a mere idea until I actually embark on the work to turn it into a concrete story.

Write the story bible.

1. Plot outline

2. Description of recurring characters

   • Background

   • Physical appearance

   • How they talk

3. Chronology

4. Other notes

The story bible serves as the standard against which all writing, by both the human and the large language model, should be made consistent.

The story bible is a set of one or more documents that functions as a reference manual for the human and is always included in the input fed to the large language model when generating scenes. It is typically a single document.

The story bible must be short enough not to overwhelm the large language model during scene generation, yet detailed enough to prevent it from forgetting details necessary to maintain consistency between scenes.

The story bible is not immutable. I anticipate the plot outline will either begin sparsely and be steadily detailed, or start as a blank slate with later developments written out as the plot evolves. If the large language model consistently introduces inconsistent details that were previously considered minor, those details should be added to the story bible.

Ideally, I would have a rough outline of the entire plot to guide the story's direction and ending. Alternatively, I might begin with everything except the plot outline, writing further plot developments as I complete earlier parts of the story. The latter approach risks a meandering story with no clear direction or conclusion, but for me, this risk is less catastrophic because I only make a story public once the entire work has been written.

To generate a scene, I write a series of prompts specifying what I expect out of the scene. A prompt is based on what is described in the plot outline in the story bible, but describes the scene in much more detail.

I specify things that I consider essential in the scene, things that happen, what the characters should say and how they say it, what the characters do and how they do it. Frequently I have to tell the large language model what not to do, as it frequently can’t help but include details which contradict the lore or have a very inappropriate tone for the scene.

One of the biggest problems the large language model has is its tendency to want to use profound-sounding expressions that are basically empty, despite instructions included to avoid overly flowery or overdramatic text. The large language model also has a lot of difficulty with using subtlety or secrecy when appropriate. Such cases require even more explicit instructions as to how to generate text.

One of the most frustrating aspects of large language models is their tendency to obsessively include extra detail that the prompt does not ask for. I often have to update prompts to tell the large language model to not include certain details, which frequently results in the large language model unable to help itself and say that certain things are not true, making the text sound like it is explaining obvious things to the readers that should not need explaining. At some point, I often find myself simply having to remove portions of the generated text manually, even if it means the remaining text feels like my typical soulless text that I am able to write myself.

Large language models, while highly adept at recognizing and filling patterns, fundamentally lack genuine understanding. They excel at bridging conceptual gaps, yet this skill appears to stem from predictive pattern completion rather than an overarching, deliberate master plan.

My creative process reserves the development of core story ideas and plots for myself. I find the narratives proposed by large language models to be often tedious in their predictability. While my plotting may not be groundbreaking, I am confident it avoids the common stereotypes found in purely AI-generated plans. I might still use generative AI as a brainstorming partner, as it can occasionally spark interesting foundational concepts, even if its complete narrative outlines often strike me as excessively dramatic and unengaging.

I rely on generative AI to handle the actual writing of prose and dialogue, which addresses a significant personal weakness. I struggle immensely with producing acceptable prose and dialogue, finding it an incredibly slow and arduous task even to reach a stereotypical level of writing.

If creative writing were my profession, I would prioritize honing my own writing skills, given the tendency of generative AI to produce conventional, stereotypical language. Alternatively, for a publishing career, I would hire a human writer to craft the prose and dialogue. However, for me, creative writing is purely a leisure pursuit, and the mechanics of writing prose and dialogue are not the enjoyable part of the hobby. Therefore, for a personal project, the timely production of serviceable, albeit stereotypical, prose and dialogue is perfectly adequate.