How to Write Prompts for AI Video Editing

In prompt-based video editing, the prompt field is the entire edit. There is no keyframe to drag, no mask to draw — the model reads your sentence and re-renders the clip accordingly. Most failed edits trace back to the prompt, not the model. This is how to write ones that work.

Rule 1: one change per prompt

A prompt describes one difference between the input and the output. That is the unit of work.

Well-scoped:

Change the car's paint to a deep metallic red.

Fighting itself:

Change the car's paint to red, remove the window reflections, add a sunset sky, and make the camera move faster.

The second prompt is four edits competing inside one generation. The model satisfies them partially and inconsistently across frames — you get a red car in some shots, an orange sky in others, and a camera move you did not want. When you need four changes, run four passes, in whatever order matters most. Each pass starts from the previous output.

Rule 2: name what must not change

Models drift when nothing anchors them. The strongest prompt pattern is a small "change / keep" pair:

Change only <the one thing>.

Keep <subject parts, background, camera movement, timing> unchanged.
Maintain natural <reflections / shadows / lighting>.

Filled in:

Change only the car's paint color to a deep metallic red.

Keep the windows, wheels, headlights, background buildings, camera
movement, and clip timing unchanged.
Maintain natural paint reflections and shadow direction.

The "keep" line is not optional politeness. It is the part that stops the wheels deforming, the logo melting, or the background smearing on frames with motion. Be specific about the parts of the subject that matter to you — if the brand logo must survive the edit, say so explicitly.

Describe outcomes, not instructions

The model re-renders pixels; it does not execute commands. Two habits follow from that:

• Target what the viewer sees. "Replace the grey sky with a clear blue sky" beats "fix the sky". "Remove the parked bicycle on the left" beats "clean up the left side".
• Keep it in prose-length. Two to four sentences is the working range. A wall of text dilutes the one change; a single vague word gives the model room to invent.

Iterate one factor at a time

When a draft is wrong, change exactly one thing in the prompt and re-run:

• the color did not take? → name the part more precisely ("the car's body panels", not "the car")
• something drifted? → add it to the "keep" line
• the edit leaked into the background? → narrow the target ("only the car, not the street")

Changing two factors at once means the next draft tells you nothing: you cannot tell which change caused the improvement. Drafts are cheap; use them to find the prompt, then switch to final quality. The workflow guide walks through the full draft → final loop.

Prompts that will not work

Be honest about the limits while writing:

• pixel-exact targets ("use exactly #E60012") — the model approximates colors; it does not match hex codes. Brand-critical color work needs post-checking in a conventional editor
• instructions about time ("cut the first two seconds") — this is an editor, not a trimmer; duration is preserved
• negation-heavy prompts ("not blue, not teal, not dark") — say what you do want; stacked negatives confuse more than they constrain
• multi-subject scoping ("change the second person from the left") — counting and ordinals are unreliable; name subjects by distinctive features instead

Where to start

The prompt library has filled-in templates for the common cases — color changes, background replacement, style transfer — each following the change/keep structure above. Open the video editor, pick the closest template, and adapt the target and the "keep" line to your clip. One change, anchored, drafted first: that is the whole method.