Ask HN: Strategies for LLMs to Generate Stories Interested in folks who have played around with this and had success in getting it to generate longer form narratives. Any strategies for keeping a narrative "on track" and remembering context you give it? |
This is the end result: https://medium.com/@robertolupi/how-beelzebub-was-born-92352...
If you like it, please upvote on HN: https://news.ycombinator.com/item?id=43599273
This is the conversation: https://g.co/gemini/share/21bbaae2a26e
I switch mid-prompt from Italian to English. It is interesting to see that Gemini reasoned in English even when writing back the initial answers in Italian.
Basically, I gave a somewhat exhaustive but generic track then asked it to give enough information to create a theatrical piece. I provided some stylistic references. A piece suitable to create a script for a tragedy is less abstract than the average output of a LLM, and forces it to create characters.
Then I asked to rewrite it as a short story a couple of time, asking to make adjustments (add more details and action, less abstract ideas) until I was satisfied.
I think I'd consider using NotebookLM or RAG-based, if I wanted something bigger and coherent.