How I use AI as a creative collaborator on “Crawford”

We are deep in pre-production on “Crawford”.

Right now, the work is all about defining the look of the show — the visual identity, the style, the tone of every frame. That means generating Master Reference Images, revising them, compiling them into shot lists, and building what will eventually become a comprehensive look book for the series. This is the phase where Crawford stops being a collection of scripts and starts becoming something you can see.

This week, I’m finishing the process of putting all twenty-four scripts through the prompt generator — the program I built to translate each script into a structured set of image prompts. Twenty-four episodes, designed to mirror a traditional broadcast season. When that’s done, I’ll have MRI prompts for every script in the series. Next week, the work shifts to revising, styling, and finalizing those images into what will be my shot list — essentially a hybrid of a traditional shot list and a storyboard.

The visual blueprint for the entire show.

But the prompt generator doesn’t do the creative thinking. We do. And that distinction matters, because it’s the core of how this process actually works.

I don’t just take the scripts and put them into the program willy-nilly and then run the prompts. I decide what I want the show to look like. That happens somewhat organically in the writing process. I’m starting to see “Crawford” in scenes. I’m starting to visualize the stories. I get an idea in my head — a feeling, a composition, a sense of how a moment should land — and then I put the scripts through the program.

But the program, before it runs, is given a creative template which Sean and I discuss, as we decide what the show is going to look like. How the characters are going to move. It’s like assigning the musical cues to the different characters in “Peter and the Wolf”. When Crawford’s on screen, he’s reliving a narrative in a standalone episode. What’s the vibe? How is he filmed? What is he wearing? That kind of stuff. Some of it’s more granular at this stage than others. Some of it comes later.

All of that goes into it, and then the script — through the program that I’ve refined, and which I continue to refine — generates a shot list prompt page. I run that through the AI of my choice — MidJourney or Kling, or whichever tool I’m using — and I get back reference images.

Those reference images, when I see them for the first time, are like a cinematographer saying to me, “Here’s an idea of how we could shoot this.”

And I am weighing that against the one I had in my own head. I will either say, “wow, your idea is so much better than mine,” or, “The way this looks is so much better than I ever imagined it.” Or I’ll say, “You know what, this is not how I thought of it. I want to have the character more in the foreground, or there should be more composition, and I want more open space here — because I wanted it to seem like he’s talking to nobody, and then when he walks over in the shot, we’re going to reveal the other characters.”

All of that represents conversations that are made feasible with the AI.

And that’s what it is — a conversation.

The same kind of conversation I’d be having with a cinematographer on set. The great thing about having the AI — and not having, say, Dariusz Khondji or Roger Deakins on set — is that I don’t have the language of a cinematographer to discuss lenses. I could do it better than the average person, but I can’t do it as well as any of those guys. But I know what I want. I know how I see it in my head, or I know the feeling I’m trying to get, and then they can help me get it. If I don’t have them, I have the AI. And the AI can help me bring that vision to life — as opposed to my own lack of a vocabulary, which would take me years to develop, stymying my creative vision.

That’s what we are doing when we make adjustments to master reference images. We’re essentially saying, “Let’s put the camera here, as opposed to here. Let’s use this kind of a lens to get this kind of a look.” The tools just allow us to have that conversation without needing a decade of technical training to participate in it.

Why You Still Can’t Skip the Human

Here’s something I discovered this week that I think is worth sharing, because it’s proof of concept for something I’ve been saying all along: you need the human in the loop.

I write a script. Let’s say it’s episode one-oh-one — a standalone Crawford episode that takes place at a rare bookstore. I run that script through the prompt generator. I get the prompts back, and they work really well. They’re consistent with the style and aesthetic and creative choices I’ve already made for the show. The style is locked. The type of camera movements and coverage is locked. Everything looks right on paper.

But once I start to generate the images from those prompts, I can’t just run through them all in sequence. Because there are multiple prompts that reference the same location — the outside of the bookstore, the inside of the bookstore — and those images need to be consistent. The bookstore has to look like the same bookstore in every shot.

Which means that before each prompt is run, a location needs to be locked. You need to know what everything looks like. You can’t go straight to a shot list without actually building out your locations first.

This, to me, is one of the places where the human interface is irreplaceable. Even if the best version of a text-to-video AI could generate minutes of footage at a time, the ability to return to a single location, have it be visually consistent, and have that work off of a single prompt — I don’t think that works. Not yet. Maybe not ever without a human making the call.

You have to build it like a traditional production. Sets have to be built. Locations have to be locked. And then the shot list is determined from those — or maybe it’s happening at the same time. But someone has to be making those decisions. Right now, that someone is me. I am currently doing the job of what would traditionally be a department head — locking locations, ensuring visual coherence, making the creative calls that hold the show together across twenty-four episodes.

And I think that’s going to remain true even as the tools get better. You’re still going to need those creative leaders. You’re probably not going to need their whole departments, because the AI is going to be able to handle a lot of the execution. That doesn’t necessarily mean those jobs disappear. It might mean those people are freed up to do other things. Or maybe it does mean there’s less of a load to carry on a single production. But here’s the upside: at a lower price point, you can make more productions. More productions means more at-bats. More at-bats means more chances for success.

The Bigger Picture

And it’s not like every creative vision has to be one that ends up in the movie theaters, either. It could be the creative vision you have for your company. Maybe you’re a single-person influencer trying to create the story of your business. You can utilize AI in this way. You can develop your own spokesperson. Why should McDonald’s be the only one to have Ronald McDonald? You can create the character — using this technology, using our process.

Or maybe it’s something more personal. Maybe you want to chronicle your family history every year. Maybe it’s a way of telling your family’s story — filming yourself and your family, capturing the year, and being able to replicate that year in and year out more feasibly because of the tools AI provides.

The technology is always going to evolve. The thing that’s not going to evolve is the DNA that goes into making a good character and telling a good story — which, frankly, go hand in hand. It’s the character informed by the world and the world informing the character.

The degree to which stories can be impactful in our lives is nearly limitless. That’s what we’re exploring with Crawford.