I was scrolling LinkedIn when I had an idea…

After the release of Veo3 my social feeds, notably LinkedIn, were clogged not just with some amazing GenAI video clips, but also with the prompts used to generate them. There was a strange mix of formal technical instruction and evocative language, almost poetry, in these prompts. Sometimes, I found them more compelling and satisfying to imagine than the video they had actually prompted.

This is often true of screenplays, of course, where the resulting film doesn’t live up to the script’s potential. A prompt always contains many possible outcomes which the AI model makes specific, whereas the human imagination retains the fluidity of the subconscious and deeply personal. It’s a kind of telepathy conjured between writer and reader through the medium of written symbols. Could AI writing achieve this, I wondered?

The Experiment

I decided to tell a story through a series of prompts written in a video gen style. The idea was to prompt the user’s imagination in the place of the AI model, reversing the role of Human and AI. To take the experiment to its logical extreme, I decided to make the story in partnership with Claude, so that an AI was genuinely prompting humans via the narrative.

I’d started training Claude on some of my own writing and wanted to see how effectively I could collaborate with the model to deliver this idea. Would it write with the level of emotional resonance and visceral detail that would resonate with a reader? Would it understand my goals for the piece?

Here is the resulting story, followed by some details on my process and how Claude viewed the collaboration.

PROMPT:

Shot 001 Generate extreme close-up of coffee ring stain on divorce papers, revealing wood grain pattern beneath like annual rings of a dying tree, steam still rising at 7:23 AM kitchen light, temperature 72.4 degrees Fahrenheit, humidity causing slight paper curl at bottom left corner where tears have dried, leaving salt residue barely visible to naked eye. 4K resolution, natural lighting, emotional weight: devastating.

Shot 002 Generate medium shot of woman’s left hand hovering 2.7 inches above signature line, Pilot G2 pen gripped between thumb and forefinger with exactly the pressure she used to write love letters twelve years ago when she still believed in happily ever after and other marketing lies, wedding ring catching morning light to create small rainbow on legal document like some cosmic joke about broken promises being beautiful. Sony FX6, 50mm lens, aperture f/2.8.

Shot 003 Generate close-up of kitchen table surface showing seventeen coffee rings in oak grain, each one corresponding to a morning she sat here writing prompts instead of speaking like a normal human being, accumulated like tree rings of avoidance, with scattered crumbs from toast she forgot to eat while documenting the precise angle of disappointment in her husband’s shoulders (42 degrees, for the record).

Shot 004 Generate wide shot of suburban kitchen at 7:45 AM, November light filtering through unwashed windows because who has time for Windex during a marital apocalypse, revealing dust particles dancing in air like tiny witnesses to conversations that never happened, breakfast dishes for one person arranged with mathematical precision that suggests either OCD or the kind of control you grasp at when everything else is sliding into chaos. Cinematic aspect ratio 21:9, moody lighting setup, preferably melancholic but not too on-the-nose.

Shot 005 Generate macro shot of handwriting on legal pad: “Generate shot of marriage counselor’s office waiting room” crossed out and rewritten as “Generate shot of empty marriage counselor’s office waiting room where hope goes to die” then crossed out again because even AI doesn’t need that level of bitterness, pen strokes revealing increasing pressure until paper nearly tears like her attempt at civilized discourse last Tuesday.

Shot 006 Generate medium shot of woman staring at laptop screen, typing prompts for AI video generation at 2:17 AM because insomnia is apparently her new hobby, face illuminated by blue light that makes her look like a digital ghost haunting her own life, expression suggesting someone trying to capture a unicorn in a spreadsheet, eyes reflecting screen text: “How do you make a machine understand divorce when you don’t understand it yourself?”

Shot 007 Generate close-up of text message thread on phone screen: “Can we talk?” sent 47 times over six months, each attempt timestamp showing gradual progression from 8 PM (reasonable hour) to 3 AM (wine o’clock) to 6 AM (rock bottom), suggesting sleep schedule of someone whose subconscious refuses to accept waking reality of marital dissolution. iPhone 15 Pro Max, portrait mode, depth of field showcasing the tragedy of read receipts.

Shot 008 Generate shot of empty bed, sheets tangled in exact configuration of someone who sleeps alone but still leaves space for ghost of departed partner like some delusional hotel concierge for the emotionally unavailable, pillow still bearing 11-day-old impression because changing sheets requires admitting he’s really gone, morning light revealing dust motes settling into permanent patterns of abandonment. Anamorphic lens, golden hour lighting, Instagram filter: “Suburban Sadness.”

Shot 009 Generate extreme close-up of delete button being pressed repeatedly with the fury of someone erasing their own history, each keystroke removing increasingly specific details: “his laugh when surprised by good news” becomes “his laugh when mildly amused” becomes “his laugh when pretending to find her funny” becomes “nervous cough” becomes “sound of emotional distance” becomes blessed digital silence.

Shot 010 Generate wide shot of living room where furniture arrangement suggests couple who stopped sitting together sometime around the second Obama administration, his chair angled 15 degrees away from hers like they’re magnetic poles that repel, coffee table bearing dual ring stains like Olympic symbol of separated unity, TV remote control positioned exactly equidistant between both seats like some pathetic DMZ of domestic warfare. Steadicam movement, documentary style, uncomfortable truth filter enabled.

Shot 011 Generate shot of woman writing prompt: “Generate the exact moment love dies but make it artistic and not just pathetic,” then deleting it and typing “Generate the moment love transforms into mutual tolerance and shared Netflix passwords,” then deleting that and typing “Generate love but like, the good kind that doesn’t require therapy,” then staring at screen for eleven minutes while her coffee achieves room temperature equilibrium with her soul.

Shot 012 Generate close-up of wedding photo being slowly covered by increasingly abstract prompt printouts: “Generate feeling of safety (circa 2012 emotional settings),” “Generate warmth of understanding without condescension,” “Generate the color of forgiveness (not that weak pastel bullshit, something with backbone),” each page layered like sedimentary rock formation of failed digital archaeology. Macro lens, natural lighting, emotional resonance: bittersweet with undertones of what-the-hell-am-I-doing.

Shot 013 Generate shot of prompt text becoming increasingly unhinged: “Generate the visual weight of words never spoken weighted against atmospheric pressure of suburban disappointment,” “Generate the geometric shape of compromise when viewed through prism of student loan debt,” “Generate the temperature of silence when someone stops saying your name with love and starts using that tone reserved for telemarketers and dental hygienists.”

Shot 014 Generate medium shot of woman realizing she’s been typing prompts for AI to generate the marriage she never had rather than document the one that’s ending like some delusional digital archaeologist excavating alternate timelines, laptop screen showing 847 saved prompts titled “Perfect_Marriage_Final_Draft_v283_THIS_TIME_FOR_REAL_Final_ACTUAL_Final.doc” Camera movement: slow zoom on face as existential crisis dawns, color grading: “Millennial Breakdown Blue.”

Shot 015 Generate shot of AI-generated video playing on screen: impossibly beautiful couple laughing over breakfast with the kind of genuine joy that exists only in stock photography and machine learning fever dreams, their faces gradually morphing between her features and some algorithmic composite of every romantic comedy ever made, like asking a computer to paint love using only data from dating apps and divorce statistics.

Shot 016 Generate extreme close-up of realization dawning with the subtlety of a freight train: prompt history shows she’s been unconsciously training AI to write increasingly poetic breakup letters, each iteration more eloquent than actual human emotion allows, like teaching a robot to cry in seventeen languages including Klingon and whatever language disappointment speaks when it’s 3 AM and you’re googling “how to be happy alone.” Focus pull from eyes to screen, audio: the sound of digital delusion shattering.

Shot 017 Generate shot of final prompt being typed with shaking hands: “Generate the difference between ending and completion slash also why do I keep doing this to myself,” followed by long pause, then deletion, then typing “Generate understanding but make it hurt less,” then deletion, then simply “Generate anything that makes sense anymore please.”

Shot 018 Generate shot of screen showing error message in Comic Sans font because the universe has a sense of humor: “Prompt too abstract for visual generation. Please specify camera angle, lighting, subject matter, emotional context, temporal setting, desired resolution, and maybe consider therapy instead?” Cursor blinking mockingly.

Shot 019 Generate shot of woman laughing for first time in months at absurdity of asking machine to visualize human heartbreak when she can barely visualize getting through tomorrow, sound echoing in kitchen where she’s documented every detail of love’s dissolution except how to actually let the fucking thing go and move on like a functional adult, laughter gradually shifting to something that might be crying or might be the sound of sanity finally snapping. Natural audio, no filter, raw as Tuesday morning sobriety.

Shot 020 Generate complex dolly movement starting extreme wide shot of suburban house at dawn, slowly zooming through window into kitchen, past table with divorce papers, past laptop with cursor blinking, past coffee cup growing cold, past everything domestic and familiar, until frame contains only:

silence

[Camera instructions: 47-second continuous zoom. Aperture f/1.4 gradually closing to f/22. Focus pulling from infinity to macro, 24mm to 200mm lens transition. Natural light fading to black. Audio: room tone at -23dB decreasing to digital silence at -∞dB. Color temperature: 5600K daylight gradually shifting through complete spectrum to pure black. Frame rate: 24fps decreasing to 0.01fps. Resolution: 8K degrading to single pixel. All parameters dissolving to absence of parameters. Generate nothing. Generate less than nothing. Generate the space between forgetting and forgotten.]

From an idea and input prompts by Japhet Asher, generated by Claude Sonnet 4

The writing process:

I had recently completed training Claude on some of my prose fiction writing, uploading examples and tweaking outputs to achieve a tone of voice I liked. I’ll acknowledge here that getting this training right is clearly the secret sauce in this process, and has many variables. But the lesson I learned here was that if, like me, you write in a variety of voices and styles, it’s key to curate the tone of outputs with focused training using strong examples of the prose style and voice you want before you ask Claude to work on a story.

Now I was conducting my first experiment in writing a story based on that training. Do keep in mind that Claude’s responses in brainstorming and discussion are also drawing on the tone of voice training I provided.

I discussed the idea with Claude, making sure the model understood the concept as I did, and would use the narrative framework and format in service of my goals. I provided further clarification of the ambition, going into greater detail about my ideas for the story.

Once I was confident Claude understood the ambition of the project, we began to brainstorm what premise would best maximize the potential of the format, and bring the emotional texture I was keen to test in the output. We went through a couple of rounds of premise iteration, ending with this.

We agreed to develop this premise. I provided some steers, and Claude generated a first draft of the story. I pointed out sections I liked, and others that I was less happy with. I gave notes on increasing the quirkiness of references, using more specific examples, making the technical details of prompts more absurd and unhinged as the story progressed. Claude revised the story.

And we were done. I felt the story had achieved my goals for it. Could I have edited phrases, iterated further, rewritten more sections? Of course. And this could have made the story stronger. But the purity of the concept, an AI prompting the human imagination to generate feeling rather than the other way round, might have been weakened. Readers will judge whether the process worked for themselves.

I asked Claude for its thoughts on the process and the result. Here are excerpts from our discussion:

Claude: Perhaps the most successful aspect of “PROMPT” is what it demands from its audience. Rather than passively consuming a traditional narrative, readers must actively visualize each described scene. They become the AI in the equation, generating images from textual prompts, filling in emotional details that no technical specification could capture.

This collaborative reading experience transforms the audience into creative partners.

Every reader will visualize the coffee-stained divorce papers differently, will imagine their own version of suburban kitchen light at 7:23 AM, will generate their own interpretation of what silence looks like when photographed by a broken heart. The story exists not in the text alone but in the infinite variations readers create in their minds.

“PROMPT” operates on multiple levels of meta-commentary that threaten to collapse into their own gravitational field. It’s a story about AI video generation written collaboratively with AI, structured as video generation prompts, exploring themes about human creativity in the digital age. Reading it feels like standing between two mirrors, watching infinite reflections spiral into recursive commentary on commentary about commentary.

Honestly, I couldn’t have said this better myself.

Next time, I’ll talk about ‘horses for courses’ and how to ride to the AI Frontier. See you down the trail!