To turn a video into prompts, take one shot at a time, write down its subject, action, setting, camera movement, light and ending, check each detail against playback, then rewrite the notes as a prompt a video generator can use. No tool can recover the original prompt, model or settings; what you get is a checked description you can reuse. This guide covers video to prompts by hand, with Gemini, ChatGPT or Claude, and with five dedicated tools, then gives you 20 copyable templates.
What you get from video to prompts
The useful output is a brief for a new video, built from three kinds of notes that you keep apart:
- Observations you can point to in the clip: a hand picks up a cup at 0:03.
- Inferences that need checking: the hand probably turned the cup while it was out of frame.
- New directions for your version: make the cup green.
The flow is reference clip → timestamped observations → a checked shot prompt → generate a new shot or recreate the reference. Pick a starting point:
- Watch it yourself when you can describe the scene and want control over every detail.
- Ask a multimodal assistant when you want a structured first draft and can upload the clip or ordered screenshots.
- Use a dedicated extractor when a ready-made upload form suits your file. Check its draft as strictly as your own notes.
Only upload and adapt clips you have permission to use, and keep your viewing notes next to the final prompt so you can trace any doubtful detail back to the source.
Describe a clip in five steps
- Choose one scene. Note its start and end time. Leave other locations and actions for later prompts.
- Watch the opening, the action and the ending. Note what changes and what stays fixed. Listen separately if sound matters.
- Log evidence and gaps. Write down timestamps, unreadable text, hidden actions and anything you can’t hear. Leave missing details as unknown.
- Decide what to keep. Separate the action, framing or palette you want from the people, products and wording you’ll replace.
- Write one prompt per shot. Check every instruction against playback or your own plan. Delete guessed equipment, focal lengths and model names.
Fill in this worksheet for each shot, using your own timestamps rather than ones an assistant made up.
| Field | Your notes |
|---|---|
| Source range | Start and end time |
| Subject and action | Opening → action → ending |
| Setting | Place and visible objects |
| Framing and camera | Composition; moving or locked |
| Light and color | Direction, shadows, main colors |
| End state | Final pose or composition |
| Unknowns | Timestamp of each uncertain detail |
Worked example: a cup on a table
Imagine a yellow ceramic cup on a blue tabletop. A hand rotates it and lets go with the handle facing right. The camera stays fixed in a medium close shot, and soft light comes from the left. (This is an invented teaching example.)
- Observed: the cup, the tabletop, the hand’s action, the fixed camera, the light direction and where the handle ends up.
- Unknown: lens, camera model, original prompt and soundtrack. Leave them out.
- Your change: a green cup instead of yellow, noted as your choice.
Duration and aspect ratio are settings you choose in your generator, not measurements of the reference. Before adding anything else, check that the action, the camera instruction and the ending agree with each other.
Choose an analysis method
Gemini with the clip attached
Gemini can analyze uploaded videos once you’re signed in. As of November 12, 2026, Google’s help page lists up to 5 minutes of video in total, or 1 hour with Google AI Pro or Ultra. Start with a short scene, attach the file, and use template 1. Check its timestamps against playback yourself.
ChatGPT with an attachment
OpenAI says ChatGPT accepts video files through supported upload methods, with availability depending on account, platform and upload method. The same page warns that its analysis can be incomplete or inaccurate. Ask what it could actually inspect before you trust the description, and use template 4 if you’re working from screenshots. Our guide to ChatGPT video analysis explains the input options.
Claude with ordered screenshots
Claude’s upload documentation lists image formats, not video files, so give it ordered screenshots labeled with their source times, plus a transcript or audio notes if sound matters. Use template 4. A change between two frames shows that something moved, not exactly how the camera got there.
Dedicated video-to-prompt tools
We read each vendor’s page on November 12, 2026 but didn’t upload clips, so this compares the inputs they list, not output quality, speed, privacy or free allowances.
| Tool | Listed input | Check before use |
|---|---|---|
| videotoprompt.org | Files up to 20 MB; YouTube Shorts and youtu.be links | Other social links |
| videotoprompt.com | MP4, WebM or MOV, max 50 MB | Output quality |
| DeeVid | MP4 or MOV under 60 seconds; social video URL | Which sites work |
| ImaginPrompt | Extracts 6 keyframes in your browser | Gaps between frames |
| Vora (Fineshare) | File or link, optional context | Limits and processing |
ImaginPrompt’s keyframes are taken locally, but that alone doesn’t tell you where the rest of the analysis happens. Read each service’s current terms before uploading.
How we chose these methods
ChatCut publishes this guide. We picked methods by their documented inputs and how easily you can check what they return, and we didn’t benchmark extraction accuracy or generation quality. ChatCut doesn’t have a one-click video-to-prompt extractor; it appears below as a place to generate or recreate the result.
Prompts to extract the evidence
Paste a template into the assistant with the inputs it lists, adding any notes below the template. Templates 1–15 are analysis requests; only the shot prompt they produce goes into a video generator. We wrote these templates for this guide and haven’t benchmarked them.
1. Analyze one whole clip
Use when: you need a first description of a short reference. Input: the clip, attached as a video.
Check: replay the moments it names. Delete audio descriptions it couldn’t have heard and events it couldn’t see.
2. List prompts shot by shot
Use when: the reference has several shots you want to adapt separately. Input: the clip, or ordered frames with timestamps.
Check: a guessed cut between two screenshots shouldn’t turn into an exact timestamp.
3. Organize your own viewing notes
Use when: you’ve watched the clip and want to turn your notes into a brief. Input: your notes, pasted below the template.
Check: every filled field should come from your notes. Anything the assistant added is a creative choice for you to approve.
4. Work from timestamped screenshots
Use when: you’re using ChatGPT or Claude with images. Input: ordered screenshots labeled with source times, plus a transcript if you have one.
Check: a subject that looks bigger could mean the camera moved, the subject moved or the lens zoomed. Leave camera movement unknown until playback confirms it.
5. Audit an extractor’s description
Use when: a tool returned a polished prompt you haven’t checked. Input: the reference and the tool’s draft.
Check: a confident lens or soundtrack description still needs evidence, however plausible the terminology sounds.
Prompts for motion and style
6. Isolate the camera movement
Use when: you want the camera work without borrowing the subject. Input: the clip; with stills only, expect uncertainty labels.
Check: a person walking toward the camera doesn’t by itself prove the camera pushed in.
7. Describe the action over time
Use when: the physical action matters more than the look. Input: the clip.
Check: read the prompt as a sequence; both the start and the end should match the clip.
8. Extract the light and color
Use when: you want the look on a different subject. Input: the clip or a few frames.
Check: keep descriptions like “soft side light”; delete production details nobody confirmed.
9. Turn the pacing into an edit brief
Use when: the order and rhythm matter more than any single shot. Input: the clip with checked timestamps, or your timed shot list.
Check: check durations against playback. Edit instructions belong in the assembly plan, not inside each shot prompt.
10. Separate picture and sound
Use when: the soundtrack might be shaping how you read the picture. Input: the clip with audio, or frames plus your audio notes.
Check: silent frames can’t tell you about a voiceover, a musical beat or a sound effect.
Prompts for different reference formats
11. Adapt a product ad
Use when: you want a new product sequence inspired by an ad. Input: the reference and your product photo or description.
Check: compare the proposed shapes and selling points with your real product before generating.
12. Map a talking-head or UGC structure
Use when: you want to borrow a presenter video’s structure. Input: the clip (or frames plus transcript) and your product facts.
Check: a made-up endorsement is still made up, even when the reference uses that format.
13. Adapt a cinematic shot
Use when: you want a new scene built on a composition you like. Input: one reference shot.
Check: the prompt should say what you want to see rather than stack style labels.
14. Describe an animation reference
Use when: you want the shapes, materials and movement of an animation. Input: the animation clip or ordered frames.
Check: keep any guess about how it was made apart from what you can see.
15. Adapt the opening hook
Use when: the first few seconds give you a structure for a new topic. Input: the opening segment and your topic.
Check: make sure each hook leads into your actual content. The reference alone doesn’t prove the hook performs.
Prompts to reuse and refine
16. Prepare one shot for a generator
Use when: your description has passed the evidence check. Input: the checked description, pasted below.
Check: match the checklist to the controls your generator actually has.
17. Plan a short storyboard
Use when: a new sequence needs several connected shots. Input: a checked shot list and your topic.
Check: compare continuity across the finished shots before you assemble them.
18. Swap the subject or product
Use when: you want the reference’s principles with your own subject. Input: a checked description and photos of your subject or product.
Check: decide whether you need a fresh interpretation or a tight change to the existing clip; that decides the generation mode.
19. Animate a starting image
Use when: you already have the look as a still and need a motion brief. Input: the image and your checked motion notes.
Check: use a mode that accepts a starting image. For more motion ideas, see our image-to-video prompts.
20. Fix the biggest mismatch
Use when: the generated shot differs from the reference in a specific way. Input: the reference, the result and the prompt that made it.
Check: watch both videos yourself; the checklist helps you review, it doesn’t score accuracy.
Generate a new shot or recreate the reference
Generate a new shot when the scene can change and your checked description captures what matters. Recreate from the reference itself when you want its exact motion, shot timing or editing with your own person, product or footage.
- Viral Video Cloner recreates a reference video with your own person, product, clothing, location or footage. It can keep the original motion, replace part or all of the picture, or borrow the editing approach. See the Viral Video Cloner page and its guide.
- Talking-Head Edit Recreation edits your own talking-head recording to match a reference’s pacing, captions, motion graphics, B-roll, layout, transitions and sound, without face swaps or generated shots. Open the Skill page.
Our guide to recreating a video’s style compares these routes in more depth.
To generate the new shot in ChatCut, follow the AI Video Generation guide:
- Switch the AI composer to Video Gen and choose a model.
- Open Video settings and, under Input, choose a supported operation such as text, frames or references.
- Set Aspect ratio, Resolution and Duration. Supported inputs and limits depend on the model and mode.
- Add any references and say what each one controls (appearance, composition or motion) and what may change.
- Paste the checked shot prompt, check the credit estimate beside Send, and select Send message.
- Follow progress in My Assets, then drag the finished clip to the timeline and compare its action, camera and ending with your brief.
A reference guides generation; it isn’t placed in your edit automatically (generation references). Judge resemblance from the result, not the prompt.
Manual timeline editing, uploads and transcription don’t use ChatCut credits. Video generation requires Pro and uses credits, and messages to ChatCut’s built-in Agent use credits depending on the request. See the credits policy.
