Skip to content

AI video prompts, with the results

Each prompt sits under its clip. Three misses sit next to follow-up attempts: garbled signs, a fold that made no sense, and a city hanging upside down. The paper rewrite simplifies the task rather than making the requested bookmark.

  • 7 prompts, 7 clips
  • 3 failures shown
  • Copy any prompt

Three failures and their rewrites

Each pair is the same idea prompted twice. The first clip of a pair is the first result; the second is the first result of the rewritten prompt.

Miss: "no text" still gave signs

The prompt ended with "no on-screen text, no logos". The shop signs are full of lettering that is not any language.

Prompt

Vertical street scene. A quiet rain-wet city street at dusk, neon reflections in the puddles, a cyclist in a yellow raincoat rides through the frame from left to right. Handheld feel, cinematic colour. No on-screen text, no logos, no watermark.

Vivix W1 fast · 5 s · 720×1264

Fix: describe a street with no shops

The rewrite describes brick walls, closed shutters and street lamps, so there is nothing to put a sign on. No lettering anywhere. The cyclist crosses the street rather than riding along it.

Prompt

Vertical street scene. A quiet rain-wet residential street at dusk between plain brick walls and closed grey shutters, lit only by warm street lamps reflected in the puddles. A cyclist in a yellow raincoat rides through the frame from left to right. Handheld feel. There are no shops, no signs, no neon and no lettering anywhere in the scene.

Vivix W1 fast · 5 s · 720×1264

Miss: a task with many steps

Asked to fold paper into a corner bookmark. The hands and paper look natural, but the folds follow no logic and the result is not a bookmark.

Prompt

Vertical top-down view. Two hands fold a sheet of blue paper into a corner bookmark on a wooden desk. Locked camera, natural daylight, every fold clearly visible. No on-screen text, no logos, no watermark.

Vivix W1 fast · 5 s · 720×1264

Simpler task: one fold

The rewrite asks for a single fold, corner to corner, then hands at rest. It made the triangle and pressed the crease, but did not attempt the original bookmark. The sheet is not an exact square.

Prompt

Vertical top-down view, locked camera, natural daylight. Two hands take a square sheet of blue paper on a wooden desk and make one single fold: corner to opposite corner, so the square becomes a triangle. The hands press the crease flat and then rest beside the triangle. No on-screen text, no logos, no watermark.

Vivix W1 fast · 5 s · 720×1264

Miss: composition left open

Asked for a skyline time-lapse in a vertical frame. The model filled the empty upper half with a second city hanging upside down.

Prompt

Vertical b-roll. Time-lapse of a city skyline at dusk as the building lights come on and clouds drift, locked-off camera. No on-screen text, no logos, no watermark.

Vivix W1 fast · 5 s · 720×1264

Fix: say what fills each part of the frame

The rewrite puts the skyline in the lower third and says the upper two thirds are open sky and nothing else. Upright city, drifting clouds, windows lighting up.

Prompt

Vertical b-roll, locked-off camera on a riverbank. A city skyline stands upright along the horizon in the lower third of the frame; the upper two thirds are open dusk sky with slow drifting clouds and nothing else. Windows light up one by one as the sky darkens. No on-screen text, no logos, no watermark.

Vivix W1 fast · 5 s · 720×1264

Recorded on 10–11 November 2026 with Vivix W1 fast through its API, at 9:16 and 720×1264. Each clip is the first result of its prompt. They are shown without sound. ClipScript does not sell this model; the people and objects in the clips are generated, not real.

What makes an AI video prompt work?

These runs suggest ideas to try, not rules proven across prompts. A single fold was simpler than the bookmark task, although it did not make a bookmark. A street of brick walls and closed shutters produced no visible signs in the second clip. A prompt placing the skyline in the lower third produced an upright city in its second clip. Each comparison also changed the requested scene, so it does not isolate one cause.

Clips on this page
7, from 7 prompts
Failed first time
3
Follow-up attempts
3; one simplified the task

One that worked first time

For comparison: a prompt with one subject, one action and its light already stated.

Honey on pancakes

Subject, action, light and speed are each one phrase, and each came through. About two seconds in, the honey gathers into a lump before it spreads.

Prompt

Vertical macro food shot. Honey drizzles slowly onto a stack of three pancakes, a little steam rising, warm morning light, slow motion. No on-screen text, no logos, no watermark.

Vivix W1 fast · 5 s · 720×1264

Recorded on 10–11 November 2026 with Vivix W1 fast through its API, at 9:16 and 720×1264. Each clip is the first result of its prompt. They are shown without sound. ClipScript does not sell this model; the people and objects in the clips are generated, not real.

A structure to write prompts in

  1. Shot

    The framing and the camera: "vertical close-up from above, locked camera". Say "locked camera" when you want no movement.

  2. Subject and the one action

    Who or what, and the single thing it does. If you wrote "then", consider a second clip.

  3. Place and light

    Describe the setting in positive terms, including the things that make unwanted details impossible.

  4. Frame split

    For vertical clips, say what occupies the top and the bottom.

What these runs do not show

  • One model

    Every clip here is Vivix W1 fast. Other models fail in other places; the three rules are habits, not laws.

  • One try each

    A prompt that failed once might pass on a second try. These are first results, not repeated trials or controlled comparisons.

  • Short clips

    All are 5 seconds. Longer clips give a model more time to drift.

AI video prompt questions

Why did "no text" fail here?

In our first run the model still drew signs on a street of shops. A different street description produced no visible lettering in the follow-up; that single pair does not establish why.

How long should a prompt be?

The prompts here run from roughly 30 to 60 words. We did not test prompt length systematically.

Should I write camera terms?

In the examples shown, the model followed some requested framing and movement. Other camera moves and repeated trials were not tested.

Can I generate clips from these prompts on ClipScript?

No. The clips on this page were recorded with Vivix W1 fast, a model ClipScript does not sell. What ClipScript does for this format: the TikTok transcript of a video you are studying, and a free voiceover script from your topic on the TikTok script generator.