Window Peeper Prompt: 5 Copy-Paste Prompts for AI Video Tools

If you already pay for a general AI video model, you do not need a dedicated window peeper site — you need a prompt that produces the five beats reliably. That is harder than it sounds, because the format depends on precise timing: a push-in that takes four seconds, a stare that holds too long, a curtain that closes the scene.

Below are the prompts we use when testing models, in both text-to-video and image-to-video forms, plus the settings and failure modes that matter.

Before the prompts: the five beats

Every prompt has to encode the same structure. If you remember one thing, remember this order:

1. Wide night exterior — an apartment block, one window lit.
2. Slow push-in — camera drifts toward the window, no cut.
3. The person eating — visible through glass, unaware.
4. The look-up — they notice the lens, freeze mid-chew, hold eye contact.
5. The curtain — pulled shut in one movement, scene ends.

Models that get 1–3 and skip 4–5 are producing atmosphere, not the trend. Push the stare explicitly, in its own sentence, with a duration.

Prompt 1: the baseline (text-to-video)

Use this on any model that accepts long prompts. It is deliberately plain — adjectives make models artistic, and this scene must not be artistic.

Vertical 9:16 video, 14 seconds. Night exterior of a plain six-storey
apartment block, shot from across a dark street. Most windows are dark;
one window on the second floor is warmly lit. The camera pushes slowly
toward that window over four seconds, no cuts, handheld but steady.
Inside, a person sits at a small table eating a burger, seen through
the glass. They look up, notice the camera, stop chewing, and stare
directly into the lens without moving for four seconds. They set the
burger down, reach up, and pull a curtain across the window. The video
ends on the closed curtain. Documentary realism, natural window light,
no music, no text.

Why it works: the camera instruction ("over four seconds, no cuts") prevents the model from cutting, which is the most common failure. The stare gets its own duration. The ending is a physical action, which models handle better than a mood.

Prompt 2: image-to-video (your photo)

This is the version that puts a specific person in the window. Upload a front-facing portrait as the reference, then use:

The person in the reference photo sits at a table in a warmly lit room
at night, seen from outside through a window. They are eating a burger.
Camera: exterior night shot of an apartment block, one lit window, slow
push-in toward the window, no cuts. The person looks up at the camera,
freezes mid-chew, and holds eye contact for four seconds. Then they pull
a curtain closed. Keep the person's face, hair and clothing from the
reference photo exactly. Vertical 9:16, 14 seconds, realistic, no text,
no music.

The phrase "keep the face, hair and clothing from the reference photo exactly" is doing the heavy lifting. Without it, image-to-video models drift toward a generic face within two seconds.

Prompt 3: the long stare (comedy cut)

If the model keeps cutting away too early, isolate the middle:

Close exterior shot through a lit window at night. A person is eating,
sees the camera, stops chewing and stares straight into the lens. The
stare holds for six seconds. Their expression does not change. Slight
handheld drift. No cut, no zoom, no reaction from the camera. The video
ends while they are still staring.

This is not the full trend, but it is the funniest six seconds of it, and it is far easier for a model to render than a five-beat arc. Some creators post this and add the curtain as a cut in the editor.

Prompt 4: the curtain ending

The hardest single beat. Models love to fade out instead of closing a curtain.

Interior, night, warm lamp light. A person sits at a table with a
half-eaten burger, looking toward the camera. They stand slightly,
reach to the right, grab a fabric curtain and pull it across the window
in one continuous movement. The frame ends filled with the curtain
fabric. No fade to black, no cut.

Say "no fade to black" explicitly. Without it you will get a dissolve, which reads as a different genre.

Prompt 5: the alternate scene

For variety, when the burger has been done to death:

Vertical 9:16, night exterior of a low apartment building. One lit
window. Slow push-in. Inside, a person in a dressing gown stands at a
fridge eating straight from a container. They notice the camera, freeze,
stare for four seconds, then pull the blind down. Realistic, handheld,
no text, no music.

Same five beats, different prop. The format survives the swap; the joke is the stare, not the food.

Negative prompts that matter

Most models accept a negative field. These are the ones worth adding:

text, subtitles, watermark, logo, cartoon, anime, oversaturated,
studio lighting, camera cut, zoom blur, smiling at camera, waving,
multiple people, extra fingers

Two of those deserve explanation:

• "smiling at camera" — models default to friendly. The joke requires a deadpan face; a smile turns it into a vlog.
• "camera cut" — without it, the model will happily cut to a new angle, which destroys the single-take illusion the format depends on.

Settings that decide the result

The seed tip is the most useful thing on this page. If a render is 90% right — good block, good stare, bad curtain — keep the seed and change one clause. Re-rolling from scratch is how you burn an afternoon.

Why prompting is still harder than a preset

Three structural problems no prompt fixes:

1. Models do not count seconds. "Four seconds" is a hint, not an instruction. You will get two, or seven.

2. Face consistency degrades. Image-to-video holds a face for a few seconds; by second ten the reference has less influence. The window gets smaller as the camera pushes in, which conveniently hides this — but not always.

3. Five beats in one generation is a lot to ask. Every beat is a chance to fail, and the failures compound. This is why dedicated tools split the work: the scene is a fixed pipeline, and your photo is the only variable.

If you want to prompt, prompt. It is genuinely fun when it lands, and a good prompt on a good model beats a preset on composition. But budget for retries — three to eight generations is normal for a clean five-beat arc.

Prompting versus a photo-to-video tool

Prompting wins on flexibility. A preset wins on reliability. If you want one clip for a group chat tonight, the preset is the rational choice; if you want a rainy window with two people and a cat, you need the prompt.

Frequently asked questions

What is the best prompt for the window peeper trend? Prompt 1 above, with the duration of the stare and the "no cuts" instruction kept intact. Those two clauses fix most failures.

Can I use a photo with a text-to-video model? Only image-to-video models accept a reference. For text-only models you are describing a person, which will not be you.

How many seconds should the stare last? Four in a 14-second clip. Six if you are isolating the stare as its own post.

Why does my render keep smiling? Because "person eating" in most training data is a pleasant scene. Add "deadpan, does not smile" and put "smiling at camera" in the negative field.

Do I need a specific model? No. The prompts are written for general models; results vary more by prompt discipline than by brand.

Can I prompt a specific apartment block? You can describe it, but you will not get a real location. For a recognisable building you have to film it.

---

If you would rather skip the retries, one photo and four minutes gets you the same five beats. If you want to understand what each beat is doing before you write your own prompt, read the trend breakdown.