Is there a prompt box?

No. The prompt is fixed per shot. You choose people and resolution.

Can I get better results by writing my own prompt?

On this site, no — there is nowhere to put it. On a general video model, a custom prompt trades recognition for flexibility.

Does the prompt use the original performer's name?

Never. Seats are described by position and instrument only.

Short answer

Tiny Desk AI does not use a user prompt. Each of the five shots has a fixed, tuned prompt that describes the camera, the blocking, the seat positions and the rule that the original performer's face, hair and hat must not carry over. A free-text prompt would fight those constraints.

Make my band video

Tiny Desk AI prompts

The prompts exist. You just never have to write one — and writing your own would make the result worse.

Updated 2026-10-29

People search for “Tiny Desk AI prompt” expecting a magic string to paste into a general video model. That is a different workflow: it gives you a clip that resembles the meme but not the meme. A fixed template removes the prompt because the template already encodes every constraint that matters.

What the fixed prompt contains

A description of the reference clip: handheld, warm practical light, record-store shelves, red keyboard in the foreground.
A keep-list: camera position and movement, framing, blocking, hand gestures, beat timing, lighting, background, instruments.
One line per seat, naming the person by position — never by clothing.
An identity line: keep that person's face, hair, skin tone, body shape and clothes from their own photo.
A discard line: do not carry over the original performer's face, hair, hat, print or logo.

Three rules that took real renders to learn

1. Never mention the mouth

Any mention of mouth, lip or lip-sync makes the model animate the mouth for the entire clip — including the laugh, where the original mouth is closed. Removing those words is what keeps the laugh reading correctly.

2. Name people by position, not by clothing

“The person in the white cap” reads as an instruction to keep the cap. “The keyboard player seated front-centre behind the red keyboard” describes a position and leaves the styling to the photo.

3. One identity per image

Using one face for two seats in the same shot makes the model replace neither, while still reporting success. That is why the studio enforces one person per seat.

If you are prompting a general model instead

Expect a looser result. General models will re-time the cuts, redraw the background and often ignore people in the back row. The reason this site renders shot by shot is that a single 19-second generation with five identities is the case that fails most often.