Building a 250,000-Word Epic with AI: The Drakenhart Experiment

A case study in AI-assisted creative production

I wanted to make a video based web series to share online. I ended up building a novel creation engine instead. Here's what happened.

The Drakenhart Saga (Brief Version)

Sera Drakenhart spent twenty-six years thinking she was ordinary. A scrappy salvage pilot, her family dead, subsisting on protein bars and a meager income from what she can salvage. Her sole companion, Prime, a GUARDIAN-class android in hiding from his makers, developed genuine sentience and feelings over three years of partnership together. His manufacturers say that's impossible. He disagrees. Then her biological mother - the Dragon Queen she'd believed was just "Aunt Aurelia" - was assassinated. Magic that is now her birthright suddenly awakens in the form of five sentient dragon tattoos on her body that talk to her and grant her powers and abilities she has no idea how to control. Her ordinary life is gone and she has to adapt quickly as forces emerge to destroy her and all she holds dear.

Together, they discover Sera can rekindle dying dimensional portals which keep magic working throughout the galaxy - a power only her bloodline has. Stakes are existential. Romance is forbidden. Politics are messy.

The Expanse meets How to Train Your Dragon, with a love story of opposites attracting, and what the nature of sentience really means.

The Video Graveyard

The original plan was to do AI-generated videos and post them as a weekly series to social media over the course of a year. Dragons, androids, romance, space portals, it had it all. I was mixing science fiction world building with fantasy elements, magic abilities, and dragons. I was going to use AI to generate a hundred or so sixty-second episodes over a long story arc exploring characters and their adventures. It was to be short form videos, TikTok style, built for discovery and social growth. Here are a couple test videos I managed to make of both Serafina Drakenhart and her companion Prime.

I spent weeks ironing out the visual looks for the various characters through image generation and prompt engineering - those dragon tattoos really took some effort to get right! I tried a few generation tools but settled on Leonardo.ai as the clear winner for their really intuitive user interface. Once I had the characters finalized, I started testing different AI video toolsets. I tried 4 in total, but LTX Studio got the most time with their Storyboard editor that generated short clips you could piece together. The core problem was consistency however. Everything could make a stunning thirty-second clip. None can make a hundred coherent episodes with the same characters looking the same way each time. Even after diving into LoRA modeling to create repeatable visuals, post production still required a level of video editing skills I simply did not possess. Video as primary output for this idea? It was now a dead end.

The visuals I created along the way weren't wasted though. My character designs and world visualizations from Leonardo became reference material for the LLM's to generate spec sheets on. This was the genesis of my pivot towards spec driven prose creation instead.

The LLM Testing Phase

Prose seemed like the answer to my video problem. I love to read, so having AI write the chapters seemed pretty cool to me. Especially since I dont consider myself all that strong at writing stories (blogs on the other hand...).

However, it soon became apparent that this opened a different set of problems to resolve.

I ran dozens of chapter generation attempts across multiple models. Claude. Gemini. Others. Each model had its own tendencies. Gemini took the story in directions Claude didnt, which meant I had to specialize into one LLM just to keep things from drifting too far apart. Claude maintained consistency better but needed more explicit constraints to avoid certain prose patterns.

The bigger issue was volume. The more words AI writes, the more words I have to read and quality check. A seven-thousand-word chapter from AI sounds efficient until you realize you're now responsible for catching every continuity error, every character acting out of voice, every plot thread that wandered off course. As I mentioned, I love to read. But this was work!

Keeping the facts of the story straight across a hundred-plus chapters is genuinely difficult. Who knows what, when? Which dragon awakened in which order? What did Sera say to Prime in chapter four that gets referenced in chapter nineteen? The human brain isn't built for this. Neither is prompting an LLM without systems to help.

Why Specifications Exist

The spec system wasn't my first instinct. It emerged from repeated failure.

Early attempts: write a prompt, get a chapter, fix what's broken, move on. This works for a few chapters. By chapter ten, you're drowning. The AI doesn't remember what it wrote before. You're manually tracking everything. The prose drifts.

So I built constraints.

I now have four tiers. The canonical worldbuilding spec holds immutable facts: how portal magic works, what factions control which sectors, character abilities and limitations. This gives the AI a reference check to make sure its not inventing something of its own.

The series arc spec ensures chapter forty-seven serves themes from chapter three. I can map out a progression of high level events from chapter 1 to whatever and have the AI check this spec to determine the proper sequencing of events. Before this was created the AI tended to rush the story and put events that should spread across an entire novel, all into a single chapter.

The chapter specification functions like test cases. Starting character states. Ending character states. Plot beats in sequence. Dialogue requirements that specify information content without dictating exact words. This spec ensured that both the AI and myself know what needs to happen, when, and where everyone lands after.

The writing style guide controls the quality of prose AI generates. Maximum three em-dashes per thousand words (jeez it loves those emdashes!). Maximum two ellipses per chapter. Strong verbs over weak ones. Show emotion through external physical reactions and work through inner dialogs rather than just stating emotions and moving on. These rules exist because AI has tendencies that compound when generating lots of words, and it loves to take shortcuts. To get the perfect writing style, I had AI re-write the same content in the prose of several different authors to figure out which styles worked best for this series. In the end, I had it mix several together which I found utterly fascinating in its own right.

The specs took time to write. But they reduce re-work and hugely increase consistency. But most of all, they make the generation process repeatable. I can now safely ask ANY of the LLM's to write a chapter of this story and it will come out relatively similar each time.

The NEW Business Model

Once I have enough chapters generated, I will open an account at Royal Road and begin sharing. Its a web based platform where serial stories like The Wandering Inn built audiences chapter by chapter.

Research shows top performers average six to nine thousand words per chapter. This has reshaped the spec even more and required more rigorous story building up-front, before the AI generates any prose for the novel.

I'm also thinking of starting a Patreon for advance chapters. Kindle Unlimited for completed seasons. Audiobooks once the story is proven. We will see as we get there. All the visual content i've already created can now pivot to become marketing material, promotional art, and supporting material.

What Actually Worked

The writing happens twice: once in the spec, once in the generation. Quality controled and reviewed by me. More time upfront, less rework downstream. Treating AI as an amplifier of my own creativity, not as an author on its own. My job is vision, structure, quality control. The AI produces words against constraints I define.

Pivoting when tools don't deliver (which has happened a lot here). Video didn't work as planned so I stopped trying to force it and moved on to something else.

Generating ten thousand words means nothing if I don't have creative control over the words and can't verify they match what I intend. The system needs to produce at a rate I can actually check.

What's Next

Season 1 is in production. Twenty chapters. Sera's awakening and her growing relationship with Prime.

I'll share updates on process, workflow, and business results as they develop.