Back to Code
Doc says we do not need roads. In my version the subtitle says codes, and a Tesla Roadster flies off instead of the DeLorean. My agent did the research, the voice test, the car and the panel. Here is how, and what each step teaches.
- Where
- In the terminal, with Claude Code
- Size
- A longer one
- Short link
- piy.show/j001
Why this Short, and why this week
Tesla showed the new Roadster in October 2026. Thrusters from SpaceX, a car that drives upside down, flips in the air and jumps over a truck. Production in 2028. Half the internet had the same thought in the same second: that is the DeLorean.
So that was the joke to make, and the week to make it. A car that lifts on thrusters, the one film ending where a car lifts on thrusters, and my own line about prompts. A remix only works while the news is warm, so speed was the brief as much as the picture.
The lesson is the one behind every Short on this site: the idea is yours, the hours are the agent's. I had the idea on the day, and a plan the same evening.
Why does a remix need to be fast? Prompt it yourself
Here is today's news and my joke about it. What is the fastest version I can post this week, and what would make it better with one more day?
One word, one car, eighteen seconds
The ending of Back to the Future. Doc says where we're going, we don't need roads. In mine the subtitle says CODES, and the DeLorean is a white Tesla Roadster. The sound is the film's own, uncut. Eighteen seconds.
I did not open an editor. I gave my agent the clip and the joke in a terminal on my Mac, and asked for a plan before a render. It came back with a page: the shots with their times, the one word to change, three ways to swap the car, and what each would cost. I picked, it built, I picked again. Nine pages later, the Short was done.
That first page is the lesson of this whole Short. A remix is a dozen small technical jobs, and the expensive mistake is to start with the hardest one. Ask for the map first. Then every step below is one prompt.
Why ask for a plan before a render? Prompt it yourself
Read this clip. List the shots with their times, the spoken words with their times, and three ways to replace the car, with the cost and the risk of each. Render nothing yet.
Find the word before you touch the picture
The joke is one word. Before anything else, the agent transcribed the clip with word timestamps on the Mac's own GPU, and found roads at 1:45.0. Then it cut the clip into shots with a scene detector and laid out a contact sheet, one frame a second, so I could point at a frame instead of describing it.
Do this for any video job. Word times tell you where the edit is, to a tenth of a second. The contact sheet turns "the bit where the car lifts" into "frame 1:48.3". The same word times came back at the end: every subtitle line appears when the word is spoken, because the cue file is the transcript.
A transcript with times is the cheapest tool in video. Ask for it first, even for a clip you know by heart.
How do I get word timestamps from a video? Prompt it yourself
Transcribe this clip with word-level timestamps. Then cut it into shots and give me a contact sheet, one frame a second, with the time under each frame.
The voice test, and why a subtitle won
The first idea was to make Doc say prompts. The agent split the dialogue from the music, cloned Doc's voice from nine seconds of his own lines, made twelve takes of the one word, and spliced each into the original at the same loudness. Then it ran a transcriber over every mix. Nine read as prompts, three read as props, and those three went.
The machine can check the word. It cannot tell you if it sounds like Doc. I listened, and it was close, but not him. So the joke moved into the subtitle. The film still says roads, the caption says CODES in red, and the gap between the two is the gag, the way subtitle parodies work.
Test the cheapest version of the risky idea first. Let the machine check what it can measure, and keep the taste call for yourself. A dropped idea that cost an hour is a good hour.
Why did the voice clone get dropped? Prompt it yourself
Separate the voice from the music. Make ten takes of this one word in the speaker's own voice, splice each into the original at the same loudness, and tell me which takes a transcriber reads as the right word.
The car, on my own Mac first
The DeLorean has to become a Roadster for about three seconds of footage, in two shots. The agent tried it on the Mac first: a mask over the car in every frame, a cut-out of the Roadster on white as the reference, and an open video model that redraws only the masked box. The first run brought the DeLorean back, a bug in the mask fill. The third run worked: a Roadster rear, the street untouched. Twenty-two minutes for 1.8 seconds, 25 GB of memory, and a soft 480p picture with wheels that never folded.
That is the figure. Left, the film. Right, the local model's best take. Free, slow, and not good enough to post.
A local run costs time, not money, so it is the right first try. It also tells you exactly what the paid tool has to beat.
How does a masked redraw keep the street? Prompt it yourself
Mask the car in every frame, use this cut-out as the reference, and redraw only the masked area with a local video model. Show me the result beside the original, frame by frame, before we try a paid model.
I can help you build yours
I build things like this with an AI agent, on my own computer. In a private session I help you do the same with your idea. Or bring it to the show, free.
Then the cloud, with one still to guide it
Two kinds of paid model were on the table. Image-to-video models get one still and invent the motion: the Roadster drove away, and the sidewalk changed. A video-to-video model gets the real shot and changes only what the prompt asks, so the street, the camera and the timing stay the film's. That one won, at about a dollar a shot.
The guide was one still, the figure here: the Roadster pasted over the DeLorean in the first frame, rough on purpose, made by an image model from the film frame and two photos. The video model blends the rough paste into a car that lifts, folds its wheels into thrusters and leaves. A small script then turned the orange thruster glow white, pixel by pixel, and left the tail lights red.
The whole car budget, every take included: about nine dollars. When the footage exists, pick the model that can see it. A model that imagines from a still is guessing the hard part.
How do I do this step myself? Prompt it yourself
Replace the car in this shot with a white Tesla Roadster. Keep the street, the camera and the timing. Start from this keyframe. One take, then show it beside the original before you spend more.
The panel: AI draws it once, code moves it
The middle of the Short is a time-circuit panel, the film's own prop. Three rows: NO CODE in the past, CODES now, PROMPTS as the destination. Fans know it in a frame.
Image models get letters wrong. So the agent generated the lit panel three times, I picked the one with every word right, and the model never touched it again. A script made the dark version from the lit one, row by row, so the two line up pixel for pixel. Then it lit the rows on a timeline: the bottom row, the middle, and the top one flickers through random letters before it locks on PROMPTS. A slow push-in and a fade from white, all from code.
The subtitles and the title are drawn by code from the real font files, never by a model, for the same reason. Letters that must be right are not the model's job. Let the model make the picture once, then let code do the motion. Code gives the same frame every time; a model gives a new one.
Why not let the model animate the panel? Prompt it yourself
Generate this panel once, with these exact words. Then write a script that makes the unlit version from the lit image and turns the rows on one by one on this timeline, so nothing else changes between frames.
Cut the picture to the sound
The sound is the film's, from the first line to the flash, with no cut. The picture had to fit it. The lift-off runs into the fly-away. The panel sits where the music goes quiet, and clears on the swell. Then the Roadster comes back down the street, over the camera, into the white flash. The end.
Behind the scenes it is one timeline in a text file, with the cues in seconds: when a subtitle shows, when a row lights, when the panel cuts. The agent renders the text as transparent layers, and the video tool stacks them over the film. Change a number, render again. Nothing was dragged by hand.
One more technical step before you post: mark it as altered or synthetic content. A film remix will most likely get a Content ID claim, which is not a strike. One clip with a clear change is the safer kind.
How do I time a cut to the music? Prompt it yourself
Keep the sound original and uncut. Fit the picture to it: list the quiet parts and the swells with their times, and put every cut on one of them.
Start with your own remix
Pick one clip and one word. Ask for the map before a render. Test the risky part locally, pay only for the shot that matters, and keep the letters in code. That is the whole method, and the first page takes an afternoon.
What do I do next? Prompt it yourself
Here is my clip and my one-word joke. Give me the map first: the shots, the words, three ways to do the hard part, and what each costs.
Do it in the terminal
I do this with Claude Code, the AI agent in the terminal. It works in a folder on your own computer, so it reads your files every time and can look things up for you.
A web chat works too, but there you paste everything in again in every new chat.
-
Get Claude Code
It comes with Claude Pro, US$20 a month, paid to Anthropic. Claude's plans
-
Make a folder for it, and open it
mkdir -p ~/projects/my-remix && cd ~/projects/my-remix
-
Start the agent
claude
-
Ask it
Here is a film clip and my joke: one word changes, and the car is a Tesla Roadster. Before you render anything, find the shots, the words with their times, and three ways to do the car, with the cost and the risk of each. Write it as a page I can read.
Never opened a terminal? In a private session I put Claude Code on your computer with you, and you start your first idea.
Prompty