I built WriteRush to help solve my writer's block, and also because I was frustrated with typical writing softwares. I wanted something that was fun, but not a gimmick; something distraction free, but still powerful. I wanted something built for writing fiction, not for writing corporate memos.
Nothing like that existed, so I built WriteRush.
The app is built in TS/Vite/React. It uses TipTap for the rich text editor, and Dexie for IndexedDB. I'm releasing a desktop version soon via Electron, and a mobile app via Capacitor. I originally built it in vanilla JS/HTML a few years ago, and have recently returned to it and rewrote the codebase.
Since I made it, it's helped me write over 150,000 words for my own projects. I'm incredibly passionate about making this the best tool in the world for writers.
I'd love to hear your feedback, or answer any questions you have!
As a writer and also the developer of an AI writing tool (Varu AI), I was really curious to see how AI would stack up against a novel I'd painstakingly wrote over a year (not published yet). So, I used Varu AI to generate a version of my story in about an hour of active work.
The AI was great in generating plot beats and character interactions, but also had some limitations in areas like worldbuilding and long-form consistency. I've detailed the good and the bad in the post.
Hi HN, Levi (poster) here. I wrote this post detailing the “plot promise” system I’m developing at Varu AI. I found traditional outlining methods limiting for the flexible, very long-form AI stories I wanted to generate. This system uses dynamic narrative threads instead.
Curious to hear thoughts on this approach vs outlines, and especially interested in discussion around the challenges mentioned, like maintaining long-range coherence with current LLM context limits. Happy to answer any questions!
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!
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.
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?