Ride Your Dog Like a Skateboard (One Prompt, the Agent Does the Rest)
Two photos and one prompt. The agent builds the still with ChatGPT Image 2, then animates it with Seedance 2.5, and hands you the finished clip. You paste once and wait.
What you need
Two photos. A clear front facing photo of your face, and a side on photo of the dog. Side on matters, because the dog is seen broadside for the entire clip and a head on dog photo gives the model nothing to build from.
Then Luma, with both photos attached to the message.
⚡ Open Luma and attach your two photos
The trick is that the prompt names the model for each half of the job. It tells the agent to build the still with ChatGPT Image 2 first, then animate that still with Seedance 2.5. You're not running two tools yourself, you're telling one agent which two to run. Leave those names out and you get whatever it picks by default, which is usually one model trying to do both jobs badly.
The prompt
Attach both photos, yourself first and the dog second, then paste this. The order matters, because the prompt refers to them by position.
Use the two images I attached. The first is the man, the second is the dog. Keep both faces exactly as they are.
Step 1, generate the still with ChatGPT Image 2. One photorealistic vertical 9:16 frame. The man stands balanced on the dog's back like it's a skateboard, both white sneakers planted flat along the dog's spine, arms out, knees bent, leaning forward as if rolling. The dog is side on to camera, low and braced, moving forward. Sunlit Mediterranean city street, pale apartment blocks, plane trees, parked bicycles, a graffiti covered concrete ledge in the foreground. Harsh midday sun, hard shadows on the pavement, deep blue sky. Low phone angle close to the ground, wide angle lens, slight barrel distortion. Real amateur skate footage, not a render. No text, no logos, no other people.
Step 2, animate that still with Seedance 2.5 as a 5 second continuous shot. The man rides the dog forward, crouches, and pops off the graffiti ledge. In the air the dog rotates one full turn around its long axis beneath his feet, legs tucked tight, ears flat. His sneakers land back down on the dog's back and they roll away down the empty street. Handheld camera running alongside at a low angle, motion blur, slight lens wobble. Real time, no slow motion, one shot, no cuts.The four things that break it
Flaring legs destroy the dog. It's tempting to ask for legs and ears splaying out mid spin, and it looks incredible in your head. The model loses track of which leg is which halfway through the rotation and hands you something with six of them. Tucked legs give it one solid shape to turn. Get a clean rotation first, then try flaring on a later pass.
Do the slow motion in your editor, not the prompt. A speed ramp asks the model to change time scale mid shot and hold everything together across the switch, and that's exactly where it falls apart. Generate at real time and ramp the flip afterwards. Takes thirty seconds in any editor and you get frame accurate control the prompt could never give you.
Watch the feet. If the sneakers detach from the dog's back in the air, the whole illusion dies instantly. When you see drift, add "his shoes never leave the dog's back" as the last line rather than rewriting anything.
One trick only. Run, pop, land. No second trick, no cutaway, no camera holding while they vanish into the distance. Every extra instruction you stack on steals attention from the rotation, and the rotation is the whole video.
Swap the dog for anything
The prompt doesn't care what animal you attach. A cat, a goat, a friend's husky, a stock photo of a golden retriever. Swap the second photo, change the word "dog", and everything else holds.
The street works the same way. Drop the Mediterranean block for a wet London high street, a Tokyo backstreet at night, an empty multi storey car park. Keep the low angle and the wide lens, because that's what makes it read as real skate footage instead of an AI clip.
For the trick itself, a kickflip is the hardest thing on the menu. If you want a cleaner result on the first try, ask for a plain ollie off the ledge with the dog's legs tucked. It lands right far more often, and it's what the original clip everybody's copying actually does.
One note worth stating plainly. This is entirely AI, no animal is involved at any point in it. Please don't go and stand on your actual dog.
Get the next one first
New prompts every week.
Free. The new drops and the tools behind them, before they hit the feed.
No spam · New issues Sunday · Unsubscribe anytime
Need it custom?
Want this built for you?
Tell me the idea and I’ll build it. An app, a tool, an automation. You don’t need to be technical.
Related guides
Prompt
One ChatGPT Prompt Turns Your Pet Into a Human
Upload one photo of your pet and this prompt turns them into a photorealistic human. Same eye color, same expression, same attitude. It should be instantly obvious which pet you started with.
Prompt
The Cinematic 'Meet Your Future Self' Video (My Exact Prompt)
The exact prompt behind the cinematic two-selves edit: your younger self and your stronger future self, lit like a boxing ring at midnight. Drop in two photos, paste this, done.
System
Clone Any Viral Reel, Frame for Frame (the AI Does the Edit)
Paste a reel link. AI reverse-engineers why it went viral, hands you a frame by frame shot list, then edits your own footage into the same video: same cuts, same text behind your head, same music hits. 🤯
Prompt
Turn One Photo of Yourself Into a Cinema-Grade Anime (Claude + Higgsfield MCP)
The exact 3-step pipeline prompt that turns a single selfie into a consistent anime protagonist, a sidekick, and a 12-second cup-final scene, no drawing, no timeline.
Frequently asked questions
No. Any clear side on photo works, including a stock photo. The model only needs one good broadside view to build from.
Because it's telling the agent which model to use for each half of the job. Image models handle two separate reference photos properly, video models tend to blend two faces into one stranger. Naming them means the still gets locked in correctly first, and Seedance 2.5 only has to move what's already right.
No. The agent calls both models for you inside Luma. You attach two photos, paste one prompt, and collect the finished clip. Open Luma here.
Almost always the legs. Make sure the prompt says legs tucked tight and never asks them to flare or splay during the rotation. If it still breaks, switch the kickflip to a plain ollie and the problem disappears.
Yes. Swap the second photo and replace the word dog in the prompt. Nothing else needs to change.