The anatomy of a useful AI video prompt

A prompt is easier to revise when you can identify what each instruction does. This worked example starts with a canvas tote on a desk, produces an actual DreamForge text prompt, then changes only the camera field. The resulting text and complete input records are available below.

We ran the same planning functions used by the free Prompt Generator on November 6, 2026. This verifies the text output and export. No video model was called, so it does not show that a tote, camera move, or fabric detail rendered successfully.

The brief: show the tote, keep the composition legible

The intended subject is an unbranded canvas tote beside a notebook. A loose handle end supplies the small action. The first version uses a static camera so that the subject movement has a clear role; the second explores an approach without changing the subject or light.

Runway’s text-to-video guide distinguishes the visible scene from its movement. That distinction informs this exercise. The field arrangement below is DreamForge’s organizational choice, rather than a required syntax for Runway or another provider. Official guidance was checked on November 6, 2026.

Enter these fields

Open the Prompt Generator, select Text-to-video and Universal, and leave the optional template unselected. Set the following values. The camera, lighting and style entries correspond to named choices in the tool; the other directions are your own text.

The subject/action fields are inside Add subject, action, and setting. The preservation, ending and duration fields are inside Keep details steady & plan the ending. Leave subject motion, scene motion, motion amount and pace at their empty or Keep my instructions values. Do not add an image for this text-only exercise.

Download the exact field records if you want to check punctuation or the full text behind a named camera choice. These JSON records are for reference; the website does not import them.

The actual first output

The planning run produced this paste-ready paragraph:

A natural-canvas tote bag stands upright on a pale wooden table beside a closed notebook. The tote has short woven handles and a plain front panel. A light breeze moves the loose end of one handle. A simple indoor desk scene. A static camera holds the composition steady. Keep the tote's outline, handle attachment points, and blank front panel consistent. Finish with both handle attachment points visible. Soft natural daylight with gentle shadows. Clean studio product photography.

The separate structure view identifies nine contributions: scene, subject, action, setting, camera, preservation, ending, lighting and visual style. The tool joins those supplied directions into sentences. It does not inspect whether the breeze is physically plausible or whether the bag would remain upright. Those are editorial decisions to resolve before rendering.

Notice where 16:9 and 5 seconds went: they appear in the settings notes, not in the paste-ready paragraph. That separation keeps a planning suggestion from being mistaken for a parameter already applied in a provider’s interface.

Change one field, inspect the difference

For the second run, change only Camera direction to Slow dolly in. The camera sentence changes from:

A static camera holds the composition steady.

to:

A slow, steady dolly moves toward the subject.

In the executed outputs, every other word in the prompt remained the same. The input comparison also found exactly one changed field: camera. The complete output record contains both results and their structure views; the readable TXT export includes both settings sheets and review checklists.

This makes the revision easier to discuss. A camera approach may draw more attention to the material and handle attachment; a static view may make the handle movement easier to isolate. Those are the reasons to test the two directions, not observations about video results we have not generated.

What this exercise established

The next production step is separate: select a provider and supported controls, render a short test, and record what actually happens. Keep the chosen model, reference assets and observations with the plan. If you need to split an action across shots, the worked storyboard example shows how purpose, timing and edit notes stay outside each shot’s prompt.