It’s sometimes hard to recall these days, but there was a time before we all relied on AI.
Some of it was utterly tiresome, like having to walk before the invention of the coach and four, add up before the invention of calculators, or talk to people in the same room before the invention of the smartphone.
But some of it, like the creative work, the craft of actually making things, turns out to be surprisingly hard to skip, even with a machine that can generate video from a text prompt.
We found this out the hard way. But once we’d worked through the initial pain both in terms of time and expense, we found it to be a genuinely exciting medium to work with.
What we made
DEPRECATED was the first. A 60-second short film we created entirely using AI video generation and AI-composed audio. It’s a found-footage horror piece. Think VHS-era camcorder footage, liminal office spaces, fluorescent lighting that feels slightly wrong.
A figure is trapped in an office that time forgot. The monitors are bulky. The carpet tiles are stained. Nothing has been updated since the mid-2000s. The whole thing sits somewhere between an analogue horror YouTube rabbit hole and a public information film about the dangers of letting your digital presence rot. It’s a marketing video for a web design studio. It doesn’t look or feel like one, and that was exactly what we wanted. [Watch DEPRECATED here.]
Having learned the process with DEPRECATED, we went back and did it again in a completely different genre. “Mr Pip Goes to Town” is a parody of 1970s British public information films, produced under the fictional “Department of Digital Vigilance.” A fox called Pip encounters a weasel representing an unscrupulous website salesperson. Same underlying message, i.e. ‘look after your website’, but a completely different tone. The fact that we could pivot from analogue horror to deadpan 1970s pastiche using the same tools and the same process tells you something about the versatility of the medium. [Watch ‘Pip Goes to Town’ here.]
“One prompt” is a fantasy
Here’s what, with cheerful naivety we hoped for as we began: describe the video, hit generate, get something usable. Here’s what actually happened: the generation process was so wildly hit-and-miss that it became clear almost immediately this could never be a one-prompt task.
Quite the opposite. We had to treat it exactly like traditional filmmaking. And the more we worked with it, the more the parallels became impossible to ignore.
In traditional production, you build the set before anyone picks up a camera. You decide on the space, dress it with props, get the details right so the environment tells its own story. With AI video, we did the same thing. Every scene description specified the room: the stained carpet tiles, the oversized CRT monitors, the strip lighting, the wood-veneer desks. Leave that to the AI’s imagination, and you’ll get a generic office. Give it the details, and you get a place that feels real, specific, and slightly wrong. That’s classic set design.
Then there’s casting. In traditional production, you choose actors based on how they look, how they move, how their presence fits the tone of the piece. We had to do the same thing, describing the character’s build, posture, clothing, and the way they move through the space. Get this wrong, and the AI recasts the character between shots. The figure in scene one looks nothing like the figure in scene three. Maintaining consistency across generations is one of the hardest technical challenges, and it mirrors exactly the continuity work that eats up time on a traditional set.
Lighting was another challenge. On a film set, the director of photography controls every light source. The colour temperature, the direction, the quality of the light, whether it’s hard or diffused, whether it flickers. We found ourselves writing the same kind of prompts for the AI. “Overhead fluorescent strip lighting, slightly green cast, one tube intermittently flickering” produces something completely different from “office lighting.”
Camera choice and lens selection matter just as much. We specified the type of camera we wanted the AI to emulate (a handheld miniDV camcorder for DEPRECATED, giving it that low-resolution, slightly unstable found-footage quality). We described the focal length, the depth of field, whether the camera was locked off on a tripod or drifting handheld. These decisions shape how the audience feels about what they’re watching, and they’re exactly the same decisions we’d make on a traditional shoot.
And then there are the takes. We scripted the entire 60 seconds in advance: every shot, every angle. We already knew what every second of the film needed to look and feel like before we generated a single frame. And even with all of that in place, we generated roughly 80 individual takes before we had enough usable footage to edit into the final film. Eighty. For 60 seconds. That’s not a typo.
Just like a traditional director calling “again” after a take that didn’t quite land, we found ourselves regenerating the same scene over and over. The framing was right, but the movement was wrong. The lighting was perfect, but the character’s hand did something impossible. The mood was exactly what we wanted for three seconds, then the AI drifted. Each generation is a take, and most takes end up on the cutting room floor.
The parallel with traditional production is uncanny. You still have a script. You still have a director’s vision. You still shoot multiple takes and select the best ones. You still sit in the edit and piece it together, cutting between angles, trimming frames, building rhythm. The camera just happens to be an algorithm instead of a lens.
The djinn problem
Occultist tomes are full of cautionary tales about summoning demons with insufficient specificity, or the perils of asking a djinn for three wishes without thinking through the exact wording. “Make me rich” might technically be fulfilled by burying you alive in a collapsing gold mine with Donald Trump.
AI video generation has exactly the same problem.
Tell it to “make a 60-second video for a web design agency”, and you’ll get something that looks like a stock footage compilation set to royalty-free panpipe music. Technically correct. Creatively bankrupt. Which is why the asking has to be meticulous.
Every prompt we wrote read more like a cinematography brief than a chatbot request. Lens type. Colour temperature. Film stock. Camera motion. The AI doesn’t ‘get the gist’. It only knows what you said.
The more detailed and prescriptive the direction, the closer the output gets to the thing in your head. Leave gaps and the AI fills them with its own interpretation. And its taste, frankly, is terrible.
The learning curve has a price tag
Here’s something nobody talks about in the breathless, pixel-perfect YouTube posts about AI video. It burns through credits at an alarming rate, especially when you’re learning.
Those 80 takes didn’t come free. Every generation costs credits, and early on, before we’d started to fine-tune our prompting approach and understood the quirks of the tool, we burned through a significant chunk of our allocation on footage that was unusable. Characters that shifted appearance mid-shot. Text that rendered as illegible smears (virtually all on-screen text in the final film is composited in post, because AI video models simply cannot render legible type). Scenes that looked gorgeous but bore no resemblance to what we’d asked for.
This is a necessary and useful learning curve. But it’s an expensive one. If you’re planning to do this, budget for the fact that your first project will cost more in credits than your fifth.
The same rules apply for sound design
We used SUNO for the soundtrack, and the same principle held. Specificity is everything. We didn’t ask for “creepy background music.” We were prescriptive about mood, instrumentation, texture. Dark ambient. Dissonant. Sporadic fragments of tuneless melody that appear without warning and stop dead. We even used it to generate individual sound elements, the buzz of a dying fluorescent tube, the hiss of static, that we layered in post alongside the main audio.
SUNO doesn’t have precise duration controls, so we specified approximate length in the prompt and trimmed to fit in the edit. Hard cuts suit the aesthetic anyway. A track that ends abruptly, like a corrupted file, is more unsettling than one that fades gracefully.
You still need the idea
This is the part that gets lost in the excitement about AI tools. The technology is genuinely remarkable. A year ago, generating video of this quality from text prompts was science fiction. But the video itself is not the hard part.
The hard part is knowing what the video should be.
No AI tool is going to tell you that a found-footage horror short is an interesting way to talk about website obsolescence. Or that a parody of a 1970s public information film, featuring a fox and a weasel, is a good way to warn businesses about dodgy web agencies. No prompt is going to generate the concept of a character trapped in a decaying office as a metaphor for digital neglect.
That’s the creative bit. The bit that still requires a human being with a point of view and the willingness to try something that might not work.
The AI is a camera. A remarkably capable, frustratingly unpredictable, occasionally brilliant camera. But a camera without a director just takes pictures of whatever’s in front of it.
What this means for your business
If you’ve watched DEPRECATED. or Mr Pip Goes to Town and recognised your own website in them (the sluggish load times, the design that made sense five years ago, the creeping feeling that your competitors have quietly overtaken you) we should talk. That’s what we do. We’ve spent 26 years building websites that actually work for businesses, not just exist on the internet.
And if you’ve watched them and thought “I want something like that for my brand,” we can do that too. We’ve offered video production as a supplement to our web work for some time, and AI has blown the doors off what’s achievable.
We’re now offering concept-to-delivery AI video production as a standalone service: scripting, direction, generation, editing, sound design, finished film. Not the “type a sentence and see what happens” kind. The kind that makes people stop scrolling.
If any of this has got you thinking, then my work here is done. Get in touch and let’s talk about how Katatomic could help your business.