How do you write deadpan for a video model?

A gaucho, a black sheep and a roof. Nearly every prompt was written once and never re-rolled, the comedy is specified down to the length of a pause, and the late prompts turn the camera operator into a character.

Original project

Listen to Meh follows a gaucho whose daily routine is repeatedly interrupted by a sheep that keeps returning to the roof of his house. It runs 5m30s, was entered in Higgsfield's AI Film Festival, and its creator made the prompts public.

Coverage: complete. 1,162 submissions, 997 unique prompts per folder and 964 across the whole project. All of them were read, split into 12 batches. I did not watch the film or download any generation, so this describes what was asked for, not what came out. When I say a change reveals a model failure, that is my inference from the correction.

In one paragraph

This film is written, not rolled. 88% of the video prompts were submitted exactly once. When a shot failed, the author changed the text and sent it again. Seedance took 97% of the credits. The prompts specify comedy the way others specify lighting: how long a pause lasts, which body part may move during a bleat, and who in the shot is not allowed to be funny.

Where the credits went:

One Seedance generation cost about 32 times one from Seedream. Multiplying each video prompt's duration by its submissions gives about 6,700 seconds requested, roughly 20 seconds for each second of the final cut.

A warning about dates. 264 prompts carry timestamps inside a single hour on the morning of 10 Sep, up to 58 of them in the same second. That looks like a bulk save and not the moment of writing, so nothing before it can be dated, and order inside that hour is unknown. Real timestamps show 211 submissions on 13 Sep and 166 on 14 Sep. The last prompt is from 14 Sep at 19:36 UTC and the final submission from 15 Sep at 11:58 UTC.

The story according to the prompts

Folder names suggest an order, but nothing in the data confirms the final cut. Several folders hold competing versions of the same idea.

The title may be the plot. The last scene states it: "the rooftop is the ONLY dry surface anywhere, and it holds every living character in the story." The sheep he spent the film removing from the roof kept going to the one place that would stay dry. He gently pets her at the end. A prompt calls it "a bright, clear, saved world after the storm". No prompt ever says the sheep knew, so that reading is mine. None of them explains how he reached the roof either.

There is an origin for the bond in the material: a sepia keepsake still of a younger gaucho bottle-feeding a newborn black lamb, which later appears as a framed photo in a rack focus to the roof. I can't tell whether it made the cut.

Writing comedy for a video model

This is the most distinctive thing in the project. There is no dialogue at all. "No dialogue." is a standing line in the constraints, and the only voice is the bleat.

The sheep is the straight man, and that is part of her design. The character sheets say where the joke lives: "The comedy comes from the total mental absence in the gaze, not from fatigue". In the mine, her nerves are in her body and never on her face: "Los nervios están en el CUERPO, no en la cara" (the nerves are in the body, not the face). Her reactions to the gaucho's efforts are comparisons: "standing planted like a boulder", "as if being carried in a parade".

Deadpan is stated as a rule, with the failure modes listed. It appears word for word in the roof shots:

THE COMEDY IS ENTIRELY DEADPAN AND COMES FROM TIMING AND FROM THE SINCERITY OF THE ATTEMPT: no cartoon proportions, no plastic skin, no mugging, no comic exaggeration, no funny faces, no sped-up motion, no cartoon sound effects.

The pause is written as an action. Instead of asking for comic timing, the prompts schedule it:

• "A PAUSE. Nothing happens."
• "He holds the point. Then he holds it a little longer than is dignified."
• "The pauses are real and are held long enough to be uncomfortable"
• "00:00 – 00:03 — NOTHING HAPPENS."
• "THE PACING IS THE POINT"

The reaction gets a register, and the register escalates. Across the chore shots the same look is rewritten step by step:

1. "a man who has heard this too many times before"
2. "weary, irritated resignation"
3. "as though the sound has personally interrupted something sacred"
4. "pure exhausted disbelief, not surprise"

The bleat is rationed and placed. It is pinned to a moment ("at second 5") or an order ("this order matters"). In the interior chores the sheep never appears: "The sheep is never shown on screen at any point — sound only, entirely off-camera." When she bleats on camera, one body part is allowed to move: "Critical: during the bleat, ONLY the sheep's mouth/jaw moves". In the mine it is cut to a single one: "Suelta un solo balido corto y bajo, casi para adentro" (a single short, low bleat, almost inward).

Scale is the joke in the containment series, measured in people. "stacking sixty humans head to toe" sets the size of a pit. The tire tower ends on "only the sheep's head is visible peering over the rubber wall".

The attempt keeps getting smaller and sadder. The roof gag is rewritten through its strategy:

• First he gestures at her, and points at the ground.
• Then he shoos her, delicately: "FROM THE DELICACY OF HIS SHOOING AGAINST HIS WEATHERED APPEARANCE".
• Finally he gives up on her entirely: "HIS GESTURES ARE COMPLAINTS, NOT COMMANDS".

Punishment arrives from the sky, and slapstick is written as fright. He laughs at the caged sheep, "risa contenida, privada y fea" (a contained, private, ugly laugh), with "nada de risa malvada de dibujo animado" (no cartoon-villain laugh). Then the storm answers. An early version has the blast throw him. A later one says "EL RAYO NO LO TOCA: lo que lo tira es el julepe." (the lightning doesn't touch him, what knocks him down is the fright). Elsewhere one sentence sets the whole tempo: "EL SUSTO ES LA ÚNICA COSA INSTANTÁNEA DEL PLANO" (the scare is the only instantaneous thing in the shot).

They argued with themselves about the hit on the head. The celebration gag went back and forth between surprise and setup. One version has "SIN NINGÚN AVISO, ALGO LE PEGA EN LA CABEZA" (without any warning, something hits him on the head). Another builds a rule of three: "Tres pasadas en escalada" (three escalating passes). The first he ignores, the second he ducks, the third connects.

How the prompts are built

There is no single template. Several coexist, and a language switch runs across them.

The [VISUAL] / [AUDIO] frame. 428 of the 542 video prompts start this way:

[VISUAL]
Single continuous close shot, 4 seconds total, deep pre-dawn night...
<<<uuid>>>: weathered older gaucho, deep-set eyes, deeply hooked nose, grey mustache
<<<uuid>>>: dim interior bedroom, iron bed frame, small window...

SHOT (single continuous take, no cuts, 4.0s total)
0.0s to 2.4s — He lies still, breathing slow and even...
2.4s — His eyelids twitch once...

CAMERA  Static extreme close-up, locked off, FOV 18°...
STYLE   60% muddy brown, 30% dark olive, 10% amber accent. WB 3200K...
CONSTRAINTS
[AUDIO]

The spec-sheet header. 94 prompts open with a stack of quality lines, then a shot description:

Style: 8K cinematic. Photorealistic — no 3D render, no game engine...
Cinematography / Lighting / Color: 60:30:10 / Camera: 180° shutter motion blur
Skin: Pore-level realism — vellus hair, asymmetric moles, capillary flush...
Acting: top-tier cinematic — micro-pauses before reactions, precise eye-line...
Physics / Composition / Continuity / Technical / Audio

SUBJECT — <<<uuid>>> (matches input 100%)
MULTISHOT —
LOCATION —
ACTION — SHOT 1 (0:00–0:05)...

The block template. About 60 prompts use SCENE CONTEXT → ACTIVE REFERENCES → LOCATION MAP → FORMAT MODE → … → POSITIVE LOCKS, with a cut list and field-of-view numbers.

The Spanish version. 48 prompts translate the header ("Estilo: 8K cinematográfico.") and add new blocks: state of the sheep, what goes where in the frame, a strict order of events, and restrictions.

Other numbers across the video prompts, in any language:

What changed when the prompts got long

From the evening of 13 Sep, the video prompts change in form. 78 are in Spanish, with a median of 9,437 characters against 4,342 for the English ones. 89 of them, in both languages, have hard line breaks in the middle of sentences at about 100 columns, which reads like text pasted out of a chat window. Spanish does not replace English: from that moment there are 78 Spanish video prompts and 71 English ones, alternating within hours, so language seems to follow which process wrote the prompt and not the date.

Here is what those prompts do that the earlier ones did not:

• The camera gets a body. "EL CAMARÓGRAFO ES UNA PERSONA MÁS DENTRO DE LA ESCENA." (the camera operator is one more person in the scene). His mistakes are specified: "Cuando el gaucho gira, él tarda en girar." (when the gaucho turns, he is late to turn). Earlier prompts describe a device. These describe someone with reflexes.
• Bans come with a physical reason. A tired man climbing out of a pit "no salta a un caballo, lo monta paso por paso" (doesn't jump onto a horse, he mounts it step by step).
• Scale is given in units. "DOCE METROS DEL SUELO" (twelve meters off the ground).
• A reference's authority is limited. "No tomes de esta referencia la luz" (don't take the light from this reference). Earlier prompts asked for the exact location, framing and lighting of a reference. Later ones say "not a fixed keyframe".
• Movement gets its own section. "NATURALIDAD DEL MOVIMIENTO" covers anticipation and dead moments, ending on "no la precisión de una animación" (not the precision of an animation).

The growth is mostly accretion. The pit chain goes from 5,981 to 20,158 characters, and most of the added text is constraints and not new content.

Fighting the model, one failure at a time

Each of these is a chain of corrections. What the model was doing is inferred from what the text forbids.

Swap the reference, keep the text. Many groups of prompts have identical text where only a reference changed: a different sheep, a different pit, a different barn. Sometimes that was the whole iteration.

Not everything converges. The look flips between "photorealistic 3D render", "no 3D render", "Hyperrealistic cinematic 3D animation" and "HYPERREALISTIC LIVE-ACTION CAPTURE". The black sheep is explicitly left out of the roof flock in some prompts and placed in it in others. The horse rides home and also stands on the roof. The header duration often disagrees with the timecodes in the text.

The darkest material carries its own hedge. The hanging and knife versions repeat lines like "without becoming an extended or exploitative depiction", and the earlier construction gags say "always reading as tender and funny rather than mistreated". Other versions replace the hanging with a tied sheep beside a pit, or with a closed, empty cage. The text gives no reason for the changes, and these versions overlap in time.

Long text and short notes

The two kinds of writing rarely mix.

Long prompts read as assistant-assisted. There is direct evidence of a chat in the loop. One prompt contains a pasted "Thought for 3s". Another opens with Spanish commentary and a code fence: "Menos negro, un rojo tenue en el horizonte" (less black, a faint red on the horizon). Some are English with Rioplatense gear names like "recado" and "cincha". No prompt states how any of this was written, so I can't say how many people were involved.

Short notes are a person at the keyboard, in Argentine Spanish, working on an attached image:

• "ponele montura al caballo" (put a saddle on the horse)
• "mas lejos", then "muchisimo mas lejos" (farther, then much, much farther)
• "TIENE QUE ESTAR FRONTAL A LA CASA" (it has to face the house head-on)

And sometimes the person breaks into the long prompt: "PORQUE CORRE COMO UN IDIOTA CARTOON?" (why does he run like a cartoon idiot?). That line, in capitals in the middle of a video prompt, is the most honest iteration note in the project.

What is worth copying

1. Write the joke's timing as events. Name the pause, give it seconds, and say what happens during it, even if the answer is nothing.
2. Give the straight man a rule. Decide who is not allowed to react, and say which body parts may move.
3. List the failure modes of the genre. "no mugging, no comic exaggeration, no funny faces" tells the model what deadpan is not.
4. Change the text between takes. Almost nobody here re-rolled. At 32 times the price of a still, each submission was treated as a new attempt.
5. Describe an operator instead of a lens. Lateness and overshoot are specific, and the model has something to play.
6. Give physical reasons with your bans, and limit what each reference is allowed to control.
7. Keep two registers. Long prompts to build, and short notes on an attached image to fix.

Still open

• Which takes entered the film. The API doesn't say, and I didn't watch it.
• The real order of events, especially which climax survived: tree and rope, pit, well, or mine and cage.
• Whether the sheep knew. It's the most satisfying reading of the title, and it is mine.
• Anything before the morning of 10 Sep, because of the bulk timestamps.