I’ve been fighting an epic struggle with my LLM friends on some writing, and failing mostly.

My mistake was trying to one-shot or two-shot the piece.

I loaded in all of the context in structured sections (3k or 5k words), ask for a plan for a piece of writing, then ask it to follow the plan and all the context and write the piece.

And every variation I tried failed. The writing was TERRIBLE! So bad.

We had a talk about it and what could be going wrong and came up with an alternative approach.

And the result is pretty good. Much better than before.

The piece is supposed to be independent research by the main character. A factual essay with a fabricated element and all filtered by her personality and and recent life events.

The story is a collection of these pieces and reading each we accrete factual knowledge of the site, knowledge of the supernatural element, knowledge of the main character and the breakdown of her life, all stepwise across the pieces.

And it should be somewhat subtle, not overly dramatic like the LLMs want to force, and not jamming in everything we know and will know.

Asking for this complex layering one shot failed, the alternative is to iteratively update the piece, as follows:

  1. Generate a “deep research” report on one aspect of the history of the site (e.g. discovery, first settlement, etc.)
  2. Write an “interesting” factual structured essay based on the report with a neutral tone.
  3. Rewrite one section of the essay to weave in a specific fabricated quote and/or supernatural element.
  4. Rewrite the entire essay by taking on the persona of the main character and weave in comments on recent life events.

This process achieves the desired effect.

Subtle main character stuff, subtle fabricated/supernatural stuff, mostly an interesting read of a factual essay.

It’s readable and does what I want and does not suck too bad. It still needs some cleaning up, but we got there.

I can also do more layering in the future, e.g. add more paranoia or obsession.

It’s obvious in retrospect and probably how one would approach the problem if writing “manually”.

I’ve read a lot on prose writing with LLMs and not seen this process called out. I’m sure others are using it. Like code, a full rewrite is “easier” for the model than a diff or list of changes.