Show HN: Varu AI – Interactively generate novel-length AI drafts | Hacker News Reader

Show HN: Varu AI – Interactively generate novel-length AI drafts

varu.us

4 pointslevihanlen9 comments

"Crown of Ash and Honor," the sample book linked on the "Start Writing" page [1], seems like a counterproductive example. With "novel-length drafts" being the main selling point for the app, this draft seems acceptable for the prose itself (it's on par with most AI writing) but doesn't really demonstrate novel-length strength.

The story begins with a boy yearning to be a knight in a war-torn fantasy world apparently without magic, then there are extra-terrestrial invaders hunting a technological artifact and he suddenly knows how to fix a shield generator by crossing wires, then it suddenly turns again into a dark fantasy story with corrupted zombie-ish warriors.

On the technical front, clicking on a link in Section 2 of the story's ToC takes me to that number in Section 2.

[1]: https://www.varu.us/books/cm9w5b2jq0001l204f2r10bnu?scene=1

I agree with your thoughts on Crown of Ash and Honor. That was made with v0.7.25, so some of the algorithms weren't as good. For example, new "plot promises" are made every 4 scenes. In the old version, the AI would gradually get more and more fantastical with these (hence why the story seemed to change genres so much). I've made it so this happens less, but it still happens a bit.

And that book was mostly me letting the AI do what it wanted with minimal guidance from me. So, I didn't edit many of the plot promises.

I'll definitely switch the sample book to "Under Fallen Banners", though. As I believe it's a bit better.

And thanks for telling me about the bug with the ToC. I'll fix that ASAP!

Hi HN,

I'm Levi Hanlen, the developer behind Varu AI: https://www.varu.us

Like many writers and readers, I had far more story ideas than time to write them all, and frankly, some stories I just wanted to read that didn't exist yet. I started building Varu AI to explore if AI could help draft these complex, long-form narratives in a more collaborative way.

Varu AI works scene-by-scene. It doesn't outline beforehand (apart from the scene outline right before each scene). The core idea is interactive guidance using what I call 'plot promises' – inspired by Brandon Sanderson's writing lectures on narrative structure. You define key plot points or character goals, and the AI works to fulfill them. If you don't like the direction, you can adjust these promises mid-stream to actively steer the story.

As a test case (which I wrote about recently), I used it to generate a 59,000-word sci-fi first draft in about 30 minutes of interaction. I'm also having it write me a 300,000 word novel (you can find this one on the website, called "Under Falling Banners". Currently still reading/generating it)

The biggest technical challenges are definitely maintaining long-range consistency of plot and managing the LLM context window effectively, while also being cost-effective. I've found that longer input prompts degrade the quality of the output. So I can't just stick in the last 50,000 words and call it a day (at least not yet). Currently, Varu AI uses techniques like dynamic scene summaries, but improving this is an ongoing effort.

It's still relatively early days – it's a public Beta. The output quality varies and absolutely produces a first draft requiring significant human editing and rewriting, not a finished, polished novel. But I'm actively innovating on the underlying systems, and I'm really excited for the future of Varu.

It's built with Next.js, TypeScript, Tailwind, Prisma, Postgres, and Stripe. Hosted on Vercel.

I'd love to get feedback from the HN community, especially from developers, writers, or anyone interested in the intersection of AI and creativity. Happy to answer any questions about the tech, the process, the challenges, or the philosophy behind it.

On that note, I'm particularly curious if others tackling long-context generation have found effective techniques beyond summarization that balance cost, speed, and quality well?

It is all about accuracy for me, as a reader, reading a book needs you to be able to forget about everything and be immersed in the world and the characters you are reading about, a simple mistake like the wrong name being used, a date being slightly wrong, or a him/her mistake can yank you out of that and really annoy you. For serious readers it does have to be 100% accuracy, 99.99% is not good enough.

A few weeks ago as a test, I copied the whole of the last draft of my last novel into two AI's and asked them to find mistakes, unanswered questions, continuity errors, plot holes etc. They found no genuine ones and lots of false ones, even after some serious prompting. The two test readers I used found five between them which would have been very embarrassing if they got through to publication.

I definitely agree.

Luckily, some things are easier for the AI to remember. I've never once come across the AI making mistakes with things like wrong character name/pronouns/etc.

The main consistency issues I've seen come from remembering the past plot, and worldbuilding.

The past plot issues come from having to store past scenes concisely. Right now, this means I can't include more than 40 of the past scene summaries. But I'm actively thinking of ways to improve this. (Note: I've found that after the input prompt surpasses a certain length, the output quality starts to degrade).

The worldbuilding issues stem from a couple things. I currently don't have any worldbuilding tracking implemented. I did in the past, but it wasn't as intuitive as storing something like character data. If I included data for locations, for example, it seemed to either not follow it at all, or it would be arbitrarily limited by it. I'll get to fixing this once I improve a few other features. In my experience, the lack of worldbuilding hasn't necessarily hindered the stories too much. Brandon Sanderson once said that out of plot, character, and setting, setting was by far the least important. I definitely agree. Though having good worldbuilding will improve the story.

Which models did you try?

Currently the model lineup includes:

- Gemini 2.0 Flash. This is what I use most often, as the quality/price ratio is amazing. (Looking into 2.5 flash preview)

- Deepseek v3 03.24

- GPT 4.1 (The most expensive model by far. The rest are very cheap. About 20x more expensive that 2.0 flash for my use cases)

- GPT 4.1 mini

- Qwen 3 235B and Qwen 3 30B MoE

- Grok 3 mini

- Gemma 3 27B

I haven’t tried any of these, recently I’ve been switching between Claude 3.7, Gemini 2.5 pro, and GPT-4.5, with each one producing something interesting once in a while. I usually sketch a plot (sci-fi short story: 10-20 pages), mention a few details of the environment, describe the main character (often based on myself), and mention relevant writing styles. Sometimes I describe how I want it to write the story, sometimes I don’t. I rarely interact with the story beyond the initial prompt, it’s easier to just resubmit if I don’t like something.

Those are some great models. The reason I don't use them is simply the cost (especially 4.5).

And that's a pretty cool process you use!

Reply on news.ycombinator.com