Write AI video prompts as a timed shot list (a viral trend breakdown)

You upload a selfie, ask for a party video, and get something that could belong to any trend. The lights look right. The face might even look right. But the gesture you were waiting for never happens, and the clip ends with someone standing around.

If your image-to-video prompt keeps failing this way, start by describing what happens on screen in order. A trend is a little sequence: a camera position, a recognizable move, a reaction and an ending. “Make it cinematic” leaves all of those decisions open.

A timed shot list gives you a way to specify them and inspect what came back. Here is how to build one, using our public Burning Bridges prompt as the worked example. Later, I'll use failures from our own tests to show where this method still needs help.

Start with the action you want viewers to recognize

Watch your reference and write down the opening pose, the changes in action and the final frame. Name visible things: a hand on a collar, friends leaning in, a laugh at the lens. “Confident” can guide the performance, but it does not tell the model where the hand goes.

Then give each beat a place in the clip. A useful planning line is:

Time range | Camera distance | Lead's action | Other people's reaction

Write the ending before adding more details to the middle. For this party scene, the payoff is the lead laughing while friends fill the frame and cheer. For a dance, it might be a held pose. That gives you something concrete to check when the render finishes.

Set the camera before describing the room

The public prompt opens with:

A vertical 9:16 party clip, exactly 10 seconds, photoreal live action in full colour, handheld phone camera close to the lead in the middle of a packed crowd. No text, no subtitles, no logos, no watermarks.

“Close to the lead” makes the face and hands the subject of the shot. “In the middle of a packed crowd” puts the camera inside the party. A distant view of a beautiful dining room would tell a different story.

The setting comes next: chandeliers, warm light, tables and green accents on the walls. Those details give the scene its look. They should support the action rather than take over the frame.

Match the duration in the tool's controls to the duration in the prompt. If your tool cannot make this length, simplify the sequence to fit the available clip. Keeping the full list and hoping the ending fits makes the result harder to diagnose.

Give the photo a clear job

This is the public prompt's “my photo” line:

The lead is the person from your reference photo, standing in the centre of the frame: their face, hair, skin tone, build and the exact clothes they wear in the photo, sharp and recognisable in every shot.

It identifies who the photo controls and which features should carry through. Attach the image as a reference in the tool; a sentence about a photo cannot supply the photo itself.

The crowd gets a separate instruction:

Every friend is a generic extra with their own face and nobody resembles a famous person.

That separates the lead from the extras. It is a request, not a guarantee against duplicated faces.

The left reference makes the face easier to see. The right hides the eyes and much of the face. This is an illustration of input choices, not a tested before-and-after result. Image created with AI.

Use a sharp photo with the face toward the camera, uncovered eyes and even light. If you care about clothing, include it in the reference. Check the framing and lighting before spending a retry.

Read the public shot list beat by beat

These are the five action lines from the public person prompt, unchanged.

0–2 seconds: establish the connection to the camera.

0-2s: close-up, the lead points finger guns at the camera, cup in the other hand, friends bouncing behind.

The opening has a camera distance, a gesture and background motion. You can inspect each separately. Did the camera stay close? Did the hand gesture happen? Did the friends move?

2–4 seconds: change the pose without changing the scene.

2-4s: rests a hand on their collar, chin up, eyes half closed, mouthing every word.

The hand, chin and eyes define the attitude. This is more specific than asking the lead to “look cool.” The lip movement is also explicit; an original soundtrack is requested elsewhere in the public prompt.

4–5.5 seconds: acknowledge the friends.

4-5.5s: turns to the friends at their side and points at them.

The attention shifts away from the lens. That prepares the crowd's response in the next beat, so the scene has an exchange rather than unrelated gestures.

5.5–7.5 seconds: let the crowd answer.

5.5-7.5s: the friends crowd in around the lead, pointing at them and hyping them up, a hand swings past the lens.

Now the extras have something to do. The hand passing the lens also gives the foreground movement. Check that this moment does not hide the lead for the rest of the clip.

7.5–10 seconds: finish with the group.

7.5-10s: the friends fill the frame around the lead, the lead covers their face laughing, then waves at the camera with the cup while everyone around throws their hands up and cheers.

The final beat contains the payoff. When reviewing, look for the laugh, the wave and the crowd response. A clip that stops during the turn has missed the ending even if the room and face look convincing.

A paragraph can work too. The advantage of this list is that a missed action is easier to locate. Timestamps express intent; they do not force the model to behave like an editing timeline.

What our tests actually got wrong

The following examples came from different trend recipes. They are useful failure patterns, not a benchmark of the public prompt above.

Content-filter rejection → cause unknown → check the permitted input and scene.

A Crazy Frog test was rejected even though the same reference photo had worked in an earlier version. The log does not establish which part triggered the filter. We cannot honestly attribute it to a particular word.

Read the tool's restrictions and use an allowed reference. For the party scene, our guide to blocked videos suggests a clear adult photo, generic extras and a simpler scene without drink-related actions. These are troubleshooting options, not a promise of acceptance.

Subject too small or too low → framing underspecified → describe the visible relationship.

In a revised Crazy Frog test, the unwanted board beneath the rider disappeared, but the rider still crouched close to the ground and became tiny in wide shots. Removing an object had not fixed the composition.

For your next test, describe how much of the subject the camera should show and what must remain visible. A short illustrative line is: “Keep the rider large in frame, with visible space between the shoes and the road.” This is a suggested change, not a separately proven fix.

Unwanted objects → a scene detail may invite them → replace that detail.

In a convenience-store test, branded drink crates appeared after the prompt described colored crates. The team suspected that the model filled in a familiar shop object. That explanation is a hypothesis, not an isolated experiment.

The proposed revision described different, plain containers. It had not been retested in that log entry. The lesson to try in your own prompt is to inspect the nouns you supplied before adding a longer list of things to avoid.

Check the opening position as well as the ending

A Toc Toc pet retest corrected the dog's ears, but the opening moves ran late. The team suspected an instruction to approach the camera had added travel before the main action. A revision made the starting position explicit; that revision had not been retested.

If the subject should already be near the lens, try a short line such as: “Start with the dancer already close to the camera.” Change that part and inspect the opening again. Do not change the photo, camera and action list all at once if you want to learn what helped.

The ending of our separate Toc Toc person test, which passed its recorded review. This is not the pet retest discussed above. Frame from an existing render. Image created with AI.

Choose how much of the process you want to control

Route	What you do	What to expect
Film it yourself	Use the shot list as directions for you and your friends	With a phone and location you already have, there is no generation fee. You still need to perform and edit the sequence.
Public prompt in a general video tool	Attach your photo, paste the public prompt and set the format	If that tool offers a suitable free tier, this is a free way to experiment. Expect watermarks, limits and retries; check its current terms.
A tool with motion-reference controls	Supply your own filmed performance or an authorized motion reference	Can specify movement beyond text. The added reference also gives you another input to inspect when something goes wrong.
A dedicated trend maker	Choose the scene and supply the required photo	Trades prompt control for a prepared workflow. Check the example, price and failed-render policy before paying.

Troubleshooting table

Symptom	Possible cause	Next thing to check
A generic party instead of the intended sequence	Scene adjectives without specific actions	Write the opening, changes in action and final frame.
Face too small to recognize	Camera distance or framing	Ask for a close view and keep the lead prominent.
The wrong person, or repeated faces	An ambiguous reference or unclear roles	Use a clear individual reference and distinguish the lead from the extras.
The scene is blocked	The reason may not be disclosed	Check permitted content and simplify allowed inputs; a retry is not guaranteed.
An invented prop breaks the scene	A noun or visual detail may invite it	Describe the intended visible state; test that change separately.
The first action starts late	Extra travel before the opening pose	State where the subject is at the first frame.
The ending never happens	Too much action for the selected duration	Remove an intermediate action and preserve the payoff.
Motion works but a pet feature drifts	Conflicting or incomplete identity cues	Check ears, coat and body shape against the photo, separately from timing.

Disclosure: I build NowStarring (nowstarring.ai), a paid, independent fan tool for making trend videos from photos, unaffiliated with the artists or creators behind these trends; it offers no free credits or trial videos.