Over the past few months, I’ve been neck-deep in large language models (LLMs) and AI, half for work, half I got curious and fell down the rabbit hole. That meant dusting off my old math brain, re-learning linear algebra, reading through transformer papers like bedtime stories, and building a bunch of scrappy demos. The goal wasn’t just to watch shiny tech demos. I wanted to know: can these tools actually help with real creative work?

To find out, I decided to build something from start to finish. Something messy and ambitious. Something that would stretch the limits of what these models can actually do.

And since I’ve always loved comic books, read a ton as a kid, even tried making some (they're probably still rotting in a box at my mom’s place), I figured, why not make one?

So I went back to my pile of old story ideas and picked one that had stuck with me: a tale set in ancient Rome, split across two timelines. I chose it on purpose, it’s a complicated story, full of historical detail, and strange twists. If an AI could keep up with this, then maybe it’s ready for more than autocomplete tricks.

The project would mix story, art, and research, just complex enough to be a real test.

The Setup #

The plan was simple in theory: use an AI to help make a full comic book from scratch. Not just a few pages or a one-off gag or a demo, an actual, coherent story with characters, scenes, dialogue, and visuals. A whole world, basically.

The story jumps between two timelines: modern-day Pompeii and the days around the eruption in 79 AD. It features a mix of English and Latin, quiet emotional beats, archaeological details, and a very specific visual style inspired by classic European comics.

I handled the story, characters, pacing, and all the writing. The AI stepped in as my assistant: helping with structure, polishing lines, and generating visual prompts to guide the artwork. It wasn’t magic, but it was surprisingly good at filling in gaps, remembering context, and keeping the tone consistent. Most days.

I organized all the sketches in a Google Sheet, added a clean layout, and overlaid the dialogue manually myself.

The Story Structure #

Once I had the story idea, the first thing I needed was structure. I didn’t want to wing it. If I wanted to evaluate how well AI could handle real creative work, I had to treat it like a real project. That meant working the way I’d approach any serious narrative: starting with a high-level story arc, then breaking it down into smaller parts.

I mapped out the story into four acts, Discovery, Contact, Fracture, and Consequence, with each act built around a core shift in the character dynamics and timeline logic. Each act contained a few scenes, and each scene had a clear purpose. Some were emotional turning points. Others were quiet. Some were purely visual. The structure was simple on paper, but layered in practice.

Once I had the acts and scenes laid out, I zoomed in. For each scene, I wrote out the key beats, the moments that had to happen. Then I started blocking those moments into pages, and those pages into panels. Usually, I’d aim for five to six panels per page. Enough to build rhythm without crowding the page. Each panel had to do real work: push the story forward, shift the mood, or say something about the characters, preferably all three.

Individual Panels #

Then came the most interesting part: the panel-level work. This is where the AI became a true creative assistant. For every panel, I fed it detailed prompts: shot size, camera angle, lighting, mood, posture, wardrobe, color palette. Everything had to match the world we were building, sun-bleached modern Pompeii in one scene, lush pre-eruption ancient Pompeii in the next. The art style was locked in early: ligne claire, with a bit of Jean Giraud sensibility. Clean lines. Natural light. Quiet surrealism. No exaggerated expressions. No comic book tropes.

The process for each panel followed a pretty clear rhythm. I’d start with a short description based on the previous step, usually a scene beat or a moment from the script. Then I’d turn that into a detailed prompt to generate a sketch.

At first, I wrote the prompts myself, long, obsessively specific, and packed with details. But pretty quickly, But soon I realized: the LLM could write the prompt for me. So I switched roles, giving it the scene and then just asking for changes. “Make it dawn.” “He should be kneeling.” “Add the slate, but don’t make it glow.”

Of course, it wasn’t always smooth. The model would forget things. A lot. Ulyses was constantly losing his bucket hat when he stepped outside. Characters from ancient Rome would randomly show up in pants. And one of my characters, Marcus, would mysteriously lose his beard in every third sketch.

When it wasn’t forgetting character details, the bigger problems were about style. Sometimes it completely ignored the color palette. Other times, it missed the composition, like the infamous struggle to make a cart caravan face the gates of Picentia. It took multiple tries to get it right.

Eventually, I learned that less was more. The best results came when I gave just enough detail to build the scene, but not so much that the prompt got overloaded. Keep it clean. Keep it focused. Let the model fill in the gaps, but make sure it remembers the hat.

The Dialog #

I wrote each scene in script format, paying close attention to what the characters said, and more importantly, what they didn’t say. The tone was dry, smart, and subtle. I wanted it to feel cinematic: no over-explaining, no fluff. Just tension, silence, and the kind of line that sticks with you because it leaves something unsaid.

I never asked the LLM to write the dialogue directly into the sketches. I tried that early on, but it didn’t work well. The placement was off, the tone was wrong, and it broke the flow. Instead, I kept it simple.

I dropped all the sketches into a Google Sheet, set up a simple layout, and added the dialogue by hand, line by line, right where it needed to go. It wasn’t fancy, but it let me control everything: pacing, placement, and tone. It was fast, flexible, and gave me full control over how the text and images worked together.

Conclusion: Did It Work? #

Partially. Using an LLM to build a full comic book was an ambitious test, and while it didn’t nail everything, it did a lot more than I expected.

As a Research Assistant #

Excellent. This was hands-down the most helpful role. The LLM saved me countless hours by quickly surfacing historical facts, translating Latin, and helping ensure consistency across the Roman world I was trying to portray. It kept the story grounded without slowing down the creative flow.

As a Sketching Assistant #

Mixed feelings. The sketching process was long and sometimes frustrating. It went like this: describe the scene, generate a prompt, tweak it, generate a sketch, give feedback, repeat, often for three or more rounds. In the end, drawing the sketches myself might have taken only 50% more time (just line work, no color), and would’ve looked better too, and I’m not even a great illustrator. Still, the process helped clarify the visual direction and made me think carefully about each panel.

As a Writing Assistant #

Surprisingly great. This was one area where the LLM really shined. Sometimes it made my writing tighter, smarter, and cleaner. I was amazed at how well it trimmed down long lines without losing meaning. It didn’t try to replace my voice; it made it sharper.

As a Director's Assistant #

Not great. The LLM often defaulted to a cinematic mindset, wide shots, fluid transitions, high-budget imagination, which doesn’t always translate well to comic book storytelling. Comics are limited in how they show time, motion, or complex shifts in setting. I kept asking myself: Will a reader actually follow this? Many times, the answer was no, and I had to simplify or restructure things manually.

It also struggled to separate what was meant for the comic from what was just part of our behind-the-scenes discussions. It often assumed readers had context they didn’t. That confusion got worse when dealing with tricky concepts like parallel timelines, bifurcated realities, or cause-and-effect across different centuries. These are hard topics even for humans, but the LLM didn’t have the tools to express them clearly in comic form.

In summary #

The LLM wasn’t a magic tool, but it was a powerful creative partner. It helped me brainstorm, rewrite, organize, research, and refine. It didn’t replace me (yet) but it pushed me to work smarter and more deliberately.

Would I do it again? Yes, but differently. I’d go in with a tighter pipeline, better tooling, and clearer boundaries between creative roles. I’d still let the LLM handle writing polish, research and sketching. The core idea, that AI can assist in real, complex storytelling, absolutely holds. And it's only going to get better.