Experiment: What does a 60K-word AI novel generated in half an hour actually look like?

Hey Reddit,

I’m Levi. Like many writers, I have far more story ideas than time to write them all. As a programmer (and someone who’s written a few unpublished books myself!), my main drive for building Varu AI actually came from wanting to read specific stories that didn’t exist yet, and knowing I couldn’t possibly write them all myself. I thought, “What if AI could help write some of these ideas, freeing me up to personally write the ones I care most deeply about?”

So, I ran an experiment to see how quickly it could generate a novel-length first draft.

The experiment

The goal was speed: could AI generate a decent novel-length draft quickly? I set up Varu AI with a basic premise (inspired by classic sci-fi tropes: a boy on a mining colony dreaming of space, escaping on a transport ship to a space academy) and let it generate scene by scene.

The process took about 30 minutes of active clicking and occasional guidance to produce 59,000 words. The core idea behind Varu AI isn’t just hitting “go”. I want to be involved in the story. So I did lots of guiding the AI with what I call “plot promises” (inspired by Brandon Sanderson’s ‘promise, progress, payoff’ concept). If I didn’t like the direction a scene was taking or a suggested plot point, I could adjust these promises to steer the narrative. For example, I prompted it to include a tournament arc at the space school and build a romance between two characters.

Okay, but was it good? (Spoiler: It’s complicated)

This is the big question. My honest answer: it depends on your definition of “good” for a first draft.

The good:

1. Surprisingly coherent: The main plot tracked logically from scene to scene.
2. Decent prose (mostly): It avoided the overly-verbose, stereotypical ChatGPT style much of the time. Some descriptions were vivid and action scenes were engaging (likely influenced by my prompts). Overall it was pretty fast paced and engaging.
3. Followed instructions: It successfully incorporated the tournament and romance subplots, weaving them in naturally.

The bad:

1. First draft issues: Plenty of plot holes and character inconsistencies popped up – standard fare for any rough draft, but probably more frequent here.
2. Uneven prose: Some sections felt bland or generic.
3. Formatting errors: About halfway through, it started generating massive paragraphs (I’ve since tweaked the system to fix this).
4. Memory limitations: Standard LLM issues exist. You can’t feed the whole preceding text back in constantly (due to cost, context window limits, and degraded output quality). My system uses scene summaries to maintain context, which mostly worked but wasn’t foolproof.

Editing

To see what it would take to polish this, I started editing. I got through about half the manuscript (roughly 30k words), in about two hours. It needed work, absolutely, but it was really fast.

Takeaways

My main takeaway is that AI like this can be a powerful tool. It generated a usable (if flawed) first draft incredibly quickly.

However, it’s not replacing human authors anytime soon. The output lacked the deeper nuance, unique voice, and careful thematic development that comes from human craft. The interactive guidance (adjusting plot promises) was crucial.

I have some genuine questions for all of you:

• What do you think this means for writers?
• How far away are we from AI writing truly compelling, publishable novels?
• What are the ethical considerations?

Looking forward to hearing your thoughts!