I generated ten premises using ChatGPT (of the form "X wants Y. But Z is in the way"), then ran them through a series of three prompts: outline, draft, revise. The outline gets the two-sentence premise and builds the outline using Holly Lisle's flash fiction method. The draft prompt gets a voice and style (derived from Dashiell Hammett's Continental Op stories by every.to's Spiral app, which includes word choice suggestions, hence all the "blooey" and "nothing stirring"). The revision prompt gets the draft and the word count, and is mostly about making sure the key to the ending is properly seeded.
Honestly, I could have done a lot better with the prompts, but optimizing the prompts felt like a bad idea, since I'd be optimizing for results on one model. So these are the prompts that Claude came up with based on my spec. Hopefully some of the stories will give you a good laugh.
The stable of models is pulled from the top twenty-ish models on OpenRouter, plus the top ten used in role-play & fiction category. Blind votes update the leaderboard using the Bradley-Terry model and Elo-style scoring (centered at 1500).
I generated ten premises using ChatGPT (of the form "X wants Y. But Z is in the way"), then ran them through a series of three prompts: outline, draft, revise. The outline gets the two-sentence premise and builds the outline using Holly Lisle's flash fiction method. The draft prompt gets a voice and style (derived from Dashiell Hammett's Continental Op stories by every.to's Spiral app, which includes word choice suggestions, hence all the "blooey" and "nothing stirring"). The revision prompt gets the draft and the word count, and is mostly about making sure the key to the ending is properly seeded.
Honestly, I could have done a lot better with the prompts, but optimizing the prompts felt like a bad idea, since I'd be optimizing for results on one model. So these are the prompts that Claude came up with based on my spec. Hopefully some of the stories will give you a good laugh.
The stable of models is pulled from the top twenty-ish models on OpenRouter, plus the top ten used in role-play & fiction category. Blind votes update the leaderboard using the Bradley-Terry model and Elo-style scoring (centered at 1500).