The Jerusalem Odyssey: a 100-page AI-generated travel memoir from a single prompt

Here's a question I found myself asking: can an AI write an entire book from a single prompt? Not a short story or a few chapters, but something approaching actual book length? I decided to find out, and the results were surprisingly impressive.

The experiment

The inspiration came from noticing Gemini 2.5's massive jump in maximum output tokens, from 8,192 to 65,536. In book terms, that's roughly the length of Sherlock Holmes or Frankenstein. I crafted a system prompt and a plot loosely based on my own emigration from Ireland to Israel, with one notable addition that didn't feature in real life: a magical, talking sloth.

The results

I tested both Gemini and Anthropic's Sonnet 3.7. Gemini produced outputs that were only about a quarter of its claimed maximum capacity. But Sonnet 3.7, which promises a staggering 128,000 tokens with specific API call formulation, produced around 27,655 words, over half of a typical novel's length. That's significant.

What impressed me most wasn't just the length but the coherence. The characters, as prompted, were held mostly consistently throughout the text. The AI independently created subplots and narratives that weren't explicitly stated in the prompt. It demonstrated a genuine ability to sustain a narrative arc over tens of thousands of words.

What this tells us

While the output falls in a literary gray zone (it's almost sort of a short book), the experiment demonstrates how far we've come in terms of continuous AI generation. The gap between claimed maximum output and actual output is still significant, but the trajectory is clear. We're moving toward a world where AI can produce genuinely book-length creative works from minimal input.

The full text, prompts, analysis data, and comparison charts are all archived on GitHub.