I am writing a story using a local large language model (LLM) on my Mac M1. I am using LM Studio to host and am trying to keep the entire novel in context as I go. This story has been in development for a year and a half, and I have migrated between several AI resources: ChatGPT, Gemini, a few local models, and Google AI Studio.
TLDR: If you want to get to the story and skip the "sausage making," look here:
My experiences have varied. In Gemini, I decided to overcome the 2000-word response limit with a continue prompt, and the Fever Dream started. Suddenly, the manuscript grew to 127,000 words, filled with meaningless hallucinations and non-sequitur paragraphs. Unapologetically.
I worked and reworked the manuscript, trying to achieve a sensible, useful outcome, but finally gave up. Last night, while watching the South Korean men's team play the gallant Chechians at a high altitude in a stadium filled with even higher expectations, I started over.
The world-building files that I had developed for the project are as follows:
01_StoryBible_World_rev
02_CharacterDossiers_name_rev
03_Structure_Outline_rev
03_Structure_MacroOutline_rev
04_Glossary_Lexicon_rev
I am using Scrivener to organize my writing, and it is both a blessing and a curse given my inexperience with this excellent program.
The rewrite sequence:
Fire up the LM Studio with the "qwen3.6-27b-heretic-uncensored-finetune-neo-code-di-imatrix-max@q5_k_m" model by DavidAU.
Load the 03_Structure_Outline and ask for literary beats by Act for the 20 chapters.
Examine the beats for quality and then
Load the 02_Character_Dossiers and ask Harry (the model's nickname) to work out the character cast for each chapter with careful attention to the King Lear allegory for the novel.
Review the assignments and output. Harry gave several more beats, each with the assigned characters and their function/role in the Chapter by Act.
Load the 04_Glossary_Lexicon file along with the prompt, "Provide a three-paragraph summary for each chapter based on the development above and considering the Glossary data provided below".
Scan the output - it's pretty good. Next prompt: "Can this story work with a third-close pov by Aria, with the narration transferred to Lyra when Aria passes? This to get agreement on the POV for the novel.
Harry agrees.
We are ready to write Chapter One.
Prompt: "Based on the above, write 2000 words for the opening chapter one".
Off to the races.
Harry started dropping connective words (the, and, etc) after about 5 chapters. We corrected. But it was getting hard, so I opened a Gemini chat and asked a few questions.
What is the reading level of this prose? Graduate level (Master's, Doctorate kind of stuff). It was not the actual words, but the length of the sentences.
I developed a prompt for Gemini to rewrite each chapter to a high school reading level, check the dialogue punctuation, and restore the connective words when they were removed.
Harry is chugging along, adding a few percent more context per chapter. This instance supports up to 262K tokens, and I have it set to around 200K.
And the result is a fever dream of loops and nonsense.
No comments:
Post a Comment