Closed on Sunday AI Prompts: Four Templates That Work

A working prompt for this trend has to do four things at once: hold one lead subject still, describe a crowd around them, imply a slow camera move, and set a light source. The four templates below are copy-ready starting points for the robed choir, a pet in the middle, a typed choir, and a text-only scene, each built so you can swap the choir without breaking the framing. They are written for Seedance 2.0 clips of 12 seconds, so every line is there to survive a short, silent render rather than a long narrative.

The trend borrows a visual grammar, not a song. The original music video arrived on 28 November 2019, and the AI remake wave picked up around 14 October 2026. What carried over is the staging: a central figure, a group arranged behind them, warm interior light, and a camera that barely moves.

A prompt for this has to describe a still image that happens to breathe. If you write an action scene, the model spends its twelve seconds trying to resolve movement and the faces soften. If you write a mood, you get something usable.

Four jobs, in order of importance:

• Name one lead. One person, one animal, one figure. Not "a group of people."
• Place the crowd. How many, how arranged, what they wear, which direction they face.
• Imply the camera. A slow push in, a gentle drift left. One move, described plainly.
• Set the light. Warm, overhead, side-lit, backlit. Light is what makes the scene read as a choir rather than a meeting.

Everything else is decoration. If a line does not serve one of those four jobs, it is costing you render time.

The baseline. Use this when you want the scene people recognise immediately.

A single figure stands centre frame in a long cream robe, facing the camera, calm expression, hands at their sides. Behind them, a choir of twelve stands in two shallow rows on a low platform, all in matching cream robes, all facing forward. Warm overhead light, soft haze, dark background. Slow push in, camera steady, no cuts. Vertical framing, subject centred, shallow depth of field.

Why it works: the lead is described before the crowd, so the model assigns detail to the right subject. "Two shallow rows" is doing real work, because deep rows turn into mush at 480p. "No cuts" is a guardrail, not a suggestion.

Swap points: robe colour, choir size, platform or floor, haze or clean air. Keep the lead sentence intact.

Same staging, different lead. This one is covered in more depth in Closed on Sunday AI with your pet, but the prompt shape is worth seeing next to the classic version.

A small dog sits centre frame on a low wooden step, facing the camera, ears up, still. Behind it, a choir of ten stands in two rows in cream robes, facing forward, slightly out of focus. Warm overhead light, soft shadow under the dog, dark background. Slow push in, camera steady, no cuts. Vertical framing, dog centred, shallow depth of field.

Two adjustments matter. First, the animal gets a stillness cue ("still", "ears up") because animals in generated video tend to drift. Second, the choir is explicitly "slightly out of focus" so the model does not try to give them equal detail and muddy the frame.

If you are uploading a photo of your own pet, the prompt describes the scene and the photo supplies the face. Keep the description of the animal generic so it does not fight the reference.

This is where the trend gets strange, and it is the template most people get wrong. A typed choir means you name the costume and the model builds the crowd from that word.

A single figure stands centre frame in a white nurse uniform, facing the camera, blank expression, arms loose. Behind them, a choir of twelve stands in three rows in matching white nurse uniforms, pale makeup, all facing forward. Cold overhead light with a warm rim, dark background, faint haze. Slow push in, camera steady, no cuts. Vertical framing, subject centred.

The failure mode is over-specifying. "Zombie nurses with bloodstained uniforms, some holding instruments, some swaying" gives the model four competing instructions and it resolves none of them well. Pick one costume word, one makeup note, one light.

Typed choirs work best when the costume is visually simple and repeated exactly. Uniformity is the joke. Variation kills it.

No reference image at all. The prompt carries everything, which means it needs a little more physical description than the others.

A single figure stands centre frame in a plain grey sweater, facing the camera, neutral expression, hands clasped. Behind them, a choir of fourteen stands in two rows in matching grey, facing forward, evenly spaced. Warm side light from the left, deep shadow on the right, dark background. Slow push in, camera steady, no cuts. Vertical framing, subject centred, shallow depth of field.

Text-only is the right choice when you have no usable photo, or when you want a look you cannot photograph. It is also the most forgiving template to iterate on, because changing one word changes the whole scene rather than fighting a reference.

Our clips are silent MP4s, so nothing in the prompt should describe sound, singing, or audio cues. Describe what the camera sees.

Across all four templates, four lines do most of the work. Get these right and the rest is polish.

• Still lead. "Facing the camera, calm expression, hands at their sides." This anchors the subject and stops the model from inventing business for them to do.
• Crowd. "A choir of twelve stands in two shallow rows, all facing forward." Number, arrangement, direction. All three, every time.
• Camera drift. "Slow push in, camera steady, no cuts." One move. The word "slow" is load-bearing.
• Light. "Warm overhead light, soft haze, dark background." A source, a quality, and a background value. Three words each, roughly.

If you find yourself rewriting a prompt and it keeps failing, check these four before anything else. Nine times out of ten the lead line drifted or the camera line grew a second move.

Most prompt problems are additions, not omissions.

Contradictions. "Warm overhead light, cool blue moonlight" gives the model two answers. "Camera steady, slow handheld drift" does the same. Pick one and delete the other.

Too many actions. "The lead turns, smiles, raises a hand, the choir sways, someone steps forward" is five actions in twelve seconds. At 480p that becomes five blurred gestures. One action maximum, and often zero is better.

Camera cuts. Any mention of a cut, a second angle, or a transition splits your twelve seconds into fragments too short to resolve. The trend is a single held shot. Write it that way.

Also worth removing: brand names, song titles, and anything describing audio. The clip is silent. Describing a soundtrack does not produce one, and it can pull the render toward a look you did not ask for.

The choir is the variable. The lead, the camera, and the light are the constants. Once you accept that, swapping choirs becomes a mechanical edit rather than a rewrite.

Change these three things and nothing else:

1. The costume word. Cream robes to white uniforms to grey knitwear. One word, applied to both the lead and the crowd so they stay in the same world.
2. The row count and number. "Two shallow rows" of twelve, or "three rows" of twelve. Keep the number stable while you test, so you know what changed.
3. The crowd's relationship to focus. "Slightly out of focus" for a pet lead, fully sharp for a robed choir where the crowd is the point.

Leave the lead sentence, the camera line, and the light line untouched. If the new choir looks wrong, you now know it was the choir and not the framing.

For more on assembling the scene around a specific idea, the choir ideas page has variations that pair with these templates, and the prompt guide covers the wider structure if you want to build your own from scratch.

One practical note on cost: a standard clip is 12 seconds, and one clip uses one lead. At 480p that is 15 credits, 720p is 30, and 1080p is 60. The smallest one-time pack is $9.90 for 15 credits, so a single 480p test is one pack. Failed renders are refunded automatically, which makes the iterate-and-check loop cheaper than it sounds. Test at 480p, confirm the framing, then re-render the winner at 1080p.

Closed on Sunday AI is an independent tool. It is not affiliated with Ye, Sunday Service, Jesus Is King, or Chick-fil-A, and it does not reproduce the original song.

Yes. Template 4 is built for that case. Templates 1 to 3 work with or without a photo; a photo supplies the face and the prompt supplies the scene.

Usually the rows are too deep or the crowd is too large for the resolution. Two shallow rows of ten to twelve holds up far better than four rows of twenty. Add "evenly spaced" and drop the row count.

No. Clips are silent MP4s. Describing singing or a soundtrack does not produce audio and can pull the render away from the framing you asked for.

One at most, and often none. A held pose with a slow camera push reads as intentional. Multiple gestures in twelve seconds read as blur.

480p at 15 credits. Confirm the lead, the crowd, and the light are right, then re-render at 720p or 1080p. Failed renders are refunded automatically, so a bad test costs you time rather than credits.

Yes. Email delivery sends a private 7-day link, and the recipient does not need an account.