How to Rewrite AI Video Prompts with GPT-6
How to Use GPT-6 to Rewrite an AI Video Prompt That Isn’t Working
The photo is fine, the prompt looks complete, but the generated video always has something off.
The person turns their head too much, the camera suddenly lurches forward, or when the hand picks up the cup, the fingers and the handle get mashed together. You add “natural movement, realistic image” at the end and try again, only for a different problem to show up somewhere else.
This is a good time to let GPT-6 help revise the prompt. But beyond the original text, you also need to tell it: what you originally wanted, what the video actually did, and which parts already look good.
The clearer the information, the more direction the revision has. In this article, GPT-6 handles the organizing and rewriting. The revised prompt still needs to go back into your video generation tool for testing.
Say what’s wrong first, then ask for a rewrite
If you just paste the prompt into GPT-6 and say “optimize this,” it might make the text fancier while keeping the exact motion that’s giving you trouble.
For example, suppose your original prompt says:
A woman sitting by a café window picks up her cup, turns her head toward the camera, smiles, and takes a sip of coffee. The camera circles around her and slowly pushes in. The image is warm and cinematic.
This prompt contains picking up the cup, turning the head, smiling, drinking coffee, plus two camera movements: orbiting and pushing in. If the video comes out wrong, just saying “it’s unnatural” won’t make it clear what to fix first.
Instead, describe the problem to GPT-6 like this:
I want to turn a café portrait photo into a short video. I only wanted the person to look toward the camera and smile slightly. But in the result, the hand holding the cup is deformed, and the camera movement is too strong. The lighting and background look good. Please revise the prompt. Focus on keeping the look toward the camera and the smile. Leave the cup on the table and use a fixed camera.
Now the problem is specific.
You don’t need to know professional terms. “The fingers and the cup handle are stuck together” is easier to understand than “the character interaction lacks realism.”
Give GPT-6 the original image and the generation settings
If the interface you’re using supports image uploads, attach the original photo too. GPT-6 Astra supports image input, so it can work with both the image and the text.
Also mention the video model you used and the clip duration you selected. That way it can revise around your actual conditions instead of making assumptions.
If the issue is a changed face or deformed hands, include a screenshot of the problem frame. If the problem happens during the motion, describe the sequence in words, like:
The first two seconds look normal, then the person suddenly turns their head sharply, and the face stops looking like the original from that point on.
A single screenshot won’t show the whole motion, so this kind of detail helps.
Here’s a template you can copy directly:
Please help me revise an image-to-video prompt.
Original prompt: [paste the text] Video model and duration: [fill in the info] What I wanted: [describe the target result] What actually went wrong: [describe the specific problem] What already looks good: [list what you want to keep]
First, suggest two possible causes, but don’t treat your guesses as certain. Then give me a revised version focused on the main issue. Keep the original idea, don’t add new actions or scenes. Finally, tell me in one sentence what you mainly changed.
Having a revision with a clear purpose makes the next comparison much easier.
Isolate the motion that keeps failing
Going back to the café example, you could first revise it to:
The person briefly looks toward the camera and gives a slight, relaxed smile. Both hands rest naturally on the table, and the cup stays beside her. The camera remains fixed, keeping the original window light and café background.
This version removes picking up the cup, drinking, and the orbiting camera. Now you can focus on whether the look toward the camera feels natural and whether the smile works.
If this version is acceptable, you can try a separate version with a slow push-in.
If picking up the cup is actually the thing you care about most, build a single shot around that motion, without adding a head turn or other changes. That makes it easier to see where the real problem lies.
Turn “more atmosphere” into a visible change
Some prompts describe a feeling without saying what the image should actually do.
For example, “make the room feel more alive” could mean curtains moving gently, sunlight shifting, a person walking in, or just a slow camera push. Pick the one you actually want to see.
Original:
Make this indoor photo come alive. Give it a cinematic feel and a natural atmosphere.
Revised:
The camera slowly moves toward the sofa. The furniture stays in place, and the sheer curtain by the window sways gently.
Product videos work the same way.
Original:
Generate a premium, attractive perfume ad.
Revised:
The perfume bottle stands upright in its original position. The camera slowly moves closer, and the front label stays facing the camera the whole time. Keep the original tabletop composition.
A description like this is easier to check: Did the camera move closer? Did the bottle spin? Did the label change?
Keep in mind that “keep the label readable” is still just a generation request, not a guarantee. You’ll still need to zoom in and check the final output yourself.
When asking GPT-6 for help, you can add:
Turn the vague atmospheric description into one visible action and one clear camera direction. Keep only the information this shot needs.
Check for contradictions in the prompt
Some sentences look fine on their own but conflict when placed together.
For example:
The camera is completely fixed, while it also orbits around the person.
Or:
The person keeps the original frontal pose, while also turning to look behind them.
In these cases, decide which requirement matters more.
If the product label must stay readable the whole time, rethink whether a full rotation makes sense. If the person needs to stay close to the original photo, try a small shift in gaze before attempting a large head turn.
You can ask GPT-6:
Check this prompt for contradictory requirements. The thing I care about most is [fill in your priority]. Revise around that and tell me what you removed.
This step helps you realize that some effects require trade-offs.
After revising, test once under the same conditions
When you paste the new prompt back into Reveedo for testing, try to keep the same photo, the same video model, the same duration, and the same other available settings.
That makes it easier to compare the two versions. If you change the photo, the model, and the prompt all at once, even a better result won’t tell you which change mattered.
First check the problem you set out to solve: Did the cup stay on the table? Did the camera stop lurching forward? When the person looks toward the camera, does the face still match the original?
Video generation has built-in randomness, so one success doesn’t mean every run will be the same. Save the version you’re happy with along with its prompt, and write a short note like:
The fixed camera helped, but the smile is still too big.
That note becomes the starting point for the next revision.
If it keeps failing, look at the photo itself
Some problems come from the source image. The hands are blocked, the label is already blurry, or the person is only shown from the shoulders up, yet the prompt asks for a full walking shot. These involve details the photo doesn’t clearly show.
If the same issue keeps appearing, try using a clearer photo, reducing the range of motion, or switching to a different video model that handles that kind of task better. Test these changes separately so you can tell which adjustment actually helped.
And when the result is already close to what you want, the next feedback only needs to address the remaining issue:
This version’s camera movement and body motion are fine, but the smile is too exaggerated. Please only adjust the expression to a closed-mouth, subtle smile. Keep everything else the same.
With feedback like that, you don’t need to explain the whole idea again. Keep what already works and keep adjusting the one detail that affects the viewing experience the most.