The Paper Machine
Script in. Vox-style paper-collage ad out. Inside Claude, on your own key.
Paste a voice-over script. Claude splits it into beats, invents one hand-cut paper metaphor per beat, pitches you the imagery before spending a cent, then generates the stills, animates each one with Veo 3.1, reads the voice-over, and stitches one vertical master with ffmpeg. No text ever appears on screen. The paper does the talking.
The machine, animated by the machine
What you get
One command chain, and you stay the director
Script to beats to stills to animated clips to voice-over to assembled MP4. You approve the imagery in chat, five small scripts do the plumbing. Nothing to install beyond Node and ffmpeg: no packages, no build step, no lockfile.
The locked paper-collage look
Torn edges, layered drop shadows, kraft-paper grain. Colour follows the emotion of each beat, not a fixed palette. No text ever appears in-frame. A hard rule, enforced in three layers.
Motion in the paper, not the camera
Things build, morph, tear, fold and assemble. Every prompt leads with a real transformation and sequences the moves. A zoom on a still is the failure this skill is built to catch.
Three worked examples, not one
A problem/solution ad, a founder origin story, and an invisible-mechanism explainer. Each one is a complete runnable set of job files. Copy the arc that matches your script and swap your subject in.
An assembler that fixes the model's bad habits
Trims every clip to its planned beat, which also slices off the random outro sting this model appends, fits the VO inside the audible threshold, ducks SFX under the voice, and normalizes the mix to -14 LUFS.
A one-command QC pass and a real setup guide
Verify builds a contact sheet and flags hot clip-tails so nothing sneaks into your master. SETUP.md takes you from nothing to a rendered example in about 15 minutes, per operating system, troubleshooting table included.
Your download should have started automatically. Tap the button if it didn't, then open SETUP.md first. You're about 15 minutes from a rendered ad.
Download the skillWhat it makes
A 16-second spot, five beats, made with nothing but this skill and one API key. Every frame and every sound below came out of the same five commands that are in the download. Turn it up, the sound effects are generated too.
Spec spot · 9:16 · 16s · unmute for the generated sound effects
Open the spot full sizeThe real Magic Spoon origin. Two guys in an empty room, a box of live crickets and a blender, a protein bar business that worked, and the cereal they actually wanted. Watch the colour: the whole frame is cool and flat until the box pours, then it goes warm and stays there. That turn is the story.
Nothing here was drawn by hand, retouched, or cut in a timeline. The stills came from one JSON file, the motion from another, and ffmpeg assembled it. Three of the five beats are keyframe pairs, meaning we drew the before and the after and the model only filled the middle. That is why the transformations land instead of drifting.
This is the b-roll bed mode: metaphor clips with their own sound effects and no voice, so you can lay your own read on top. The skill also generates the voice-over and mixes it in for you. Same job files, one more command.
- 5 beats
- 5 stills
- 5 clips
- 0 words on screen
- -14 LUFS
The look holds across categories. Three more, one still each, same skill, same style block:
Spec work. These brands are not clients and had no involvement. We picked stories people already know so you can judge the format instead of the writing.
How it works
- Claude splits your script into beats and invents one paper metaphor per beat, then pitches you the full imagery table and waits. Nothing generates until you nod. This gate is the whole craft.
- Stills first, always. One per beat, a few cents each. You review them with your own eyes before a dollar goes to video.
- Veo 3.1 animates each still image-to-video, with the bans (no text, no music, no speech) loaded at the top of the prompt, because video models ignore rules buried at the bottom.
- The voice-over is read one line per beat, each matched to that beat's emotion. Tense for the problem, warm for the payoff. A flat read is why most AI ads feel dead.
- ffmpeg assembles the master: trims each clip to its beat, fits the VO, ducks the sound effects under the voice, normalizes to feed loudness. Then one verify command builds a contact sheet so you QC the whole ad in a single look.
What running it costs
Fair warning up front
This won't replace your studio work or your UGC pipeline, and it isn't trying to. It makes one specific thing extremely well: the editorial, metaphor-driven explainer your account probably doesn't have. The format that stops a scroll precisely because it doesn't look like an ad.
Is it perfect? No. Video models miss, and this skill's whole architecture is about catching the misses cheaply: stills before clips, one clip before five, a contact sheet at the end.
Will every ad it makes win? No. Nothing does. It's fifteen minutes and a few dollars to put a genuinely different format into your next test, and that's usually worth a slot.
Do I need to know code? No. Claude runs the commands. You approve pictures and make creative calls.
Do I use your API key? No, and you can't. You bring your own, you see every cent on your own dashboard, and nothing you make ever touches us.
This is one format. A real Meta account burns through a dozen a month.
That is the part a skill cannot do for you. We build AI performance creative for DTC brands at the volume the algorithm actually needs, and we read the results back into the next batch. If the output on this page is the bar you want and you would rather not run the machine yourself, take fifteen minutes with us.
Book 15 minutes with LeoNO DECK, NO PITCH SEQUENCE. WE PULL YOUR AD ACCOUNT
AND TALK ABOUT WHAT IS MISSING FROM IT.
More free work
THE PAPER MACHINE IS FREE TO USE, SHARE AND REMIX. MODEL COSTS BILL TO YOUR OWN API KEY.