How to Make AI Videos with ChatGPT or Claude

ChatGPT and Claude can plan an AI video, but they don’t render one. The workflow that works is to let the assistant write the script and a numbered list of shot prompts, then paste each shot into a video model such as Kling 4.0, Seedance 2.5, MiniMax H3, Veo 3.1 or Wan 3.0 on VdoBloom. You generate each clip, then string the clips together. Below is a copy-paste “director prompt” that makes the assistant output shot prompts ready to paste, plus two worked examples.

Updated October 2026.

Why the assistant can’t make the video itself

A chat assistant is a language model. It reads text (and images) and writes text back. Video generation is a separate kind of model, trained on footage, that turns a written description into moving frames. Some chat products bolt on outside apps or connectors that call a video model for you. Even then the assistant only writes the request, and a separate video model does the rendering.

So the useful split is:

The assistant writes the hook, the script, the shot list and one detailed prompt per shot. It’s fast at this and easy to revise (“make shot 3 more dramatic”, “cut it to 15 seconds”).
The video model renders each shot as a short clip, usually 5 to 10 seconds, sometimes longer.
You pick the model per shot, check the results, regenerate the weak ones and assemble the final cut.

VdoBloom has no ChatGPT app or Claude connector. You copy the prompts out of the chat and paste them into the VdoBloom tools. It takes a few seconds per shot, and you get to see and approve the cost of every clip before it runs.

What makes a good shot prompt

Video models respond best to short, concrete prompts that cover six things:

Subject: who or what is on screen, described visually.
Action: one clear movement. Models struggle with three things happening at once.
Setting: place, time of day, weather.
Camera: shot size (close-up, wide) and one move (slow push-in, orbit, handheld, static).
Light and look: golden hour, soft window light, neon, film grain.
Duration and format: 5 or 10 seconds, 9:16 for Reels and TikTok, 16:9 for YouTube.

Assistants tend to write long, flowery paragraphs. The template below forces them into this structure instead.

The director prompt (copy and paste)

Paste this into ChatGPT or Claude and fill in the brackets:

You are a film director planning a short AI-generated video.
Each shot will be rendered separately by an AI video model,
so every shot prompt must stand on its own.

Video idea: [describe the video]
Platform and aspect ratio: [e.g. Instagram Reel, 9:16]
Total length: [e.g. 20 seconds]
Style: [e.g. bright, clean, premium product ad]

Output:
1. A one-line hook.
2. A numbered shot list. For each shot give:
   - Duration (5 or 10 seconds)
   - Prompt: 40-70 words, present tense, covering subject,
     one action, setting, camera shot size and one camera move,
     lighting and look. Repeat key visual details (colours,
     clothing, product shape) in every shot so they stay consistent.
   - Start-frame idea: one sentence describing a still image
     that could be the first frame.
   - Optional voiceover or on-screen text.
Rules: one action per shot, no text or logos inside the
video frame, no camera cuts inside a shot.

The last rules matter. Models often garble lettering, so add titles and captions in an editor afterwards. And because each shot is generated on its own, repeating the visual details in every prompt (the red jacket, the matte black bottle) is what keeps the clips looking like one video.

Worked example 1: a 20-second product ad

Input to the assistant: “Video idea: launch ad for a matte black insulated water bottle with a bamboo lid. Instagram Reel, 9:16, 20 seconds, premium and minimal.”

A typical response (trimmed):

Hook: "The last bottle you'll buy."

Shot 1 - 5s
Prompt: Close-up of a matte black insulated water bottle
with a light bamboo lid standing on a wet slate surface,
water droplets beading on the bottle, slow push-in camera,
soft overhead studio light with a cool blue rim light,
dark minimal background, premium commercial look.

Shot 2 - 5s
Prompt: A hand twists the bamboo lid off the matte black
bottle, a thin wisp of cold vapour rises, macro close-up,
static camera, crisp side light, shallow depth of field.

Shot 3 - 5s
Prompt: A hiker in a grey fleece drinks from the matte black
bottle with bamboo lid on a misty mountain ridge at sunrise,
medium shot, slow orbit around the hiker, warm golden light.

Shot 4 - 5s
Prompt: The matte black bottle with bamboo lid rotates slowly
on a turntable, studio black background, soft top light,
gentle reflections, smooth 360 product spin.

How to render it on VdoBloom:

Open text-to-video, set 9:16 and 5 seconds.
Paste shot 1 and pick a model. For glossy product close-ups, Kling 4.0 or Veo 3.1 are good first choices.
The credit cost shows before you generate. Run it, check the clip, and regenerate if the bottle shape drifts.
Repeat for shots 2–4. Shot 3 has a person moving through a landscape, which suits Seedance 2.5 or MiniMax H3.
Download the clips and put them in order in any editor, adding the hook as on-screen text.

If your product has to look exactly like the real one, skip text-to-video for the product shots and use the image-first workflow below with a photo of the actual bottle.

Worked example 2: a 30-second travel reel

Input: “Video idea: a dreamy travel reel of Kyoto in autumn, 9:16, 30 seconds, warm cinematic film look, no people facing the camera.”

Shot 1 - 5s
Prompt: Wide shot of a quiet stone path lined with red and
orange maple trees in Kyoto, leaves drifting down, slow
forward dolly, soft morning haze, warm film grain.

Shot 2 - 5s
Prompt: Close-up of a single red maple leaf landing on dark
still pond water, ripples spreading, static overhead camera,
soft diffused light, warm cinematic colour.

Shot 3 - 10s
Prompt: A woman in a beige coat walks away from the camera
under a row of vermilion torii gates, medium-wide shot,
slow follow camera behind her, dappled autumn sunlight.

Shot 4 - 5s
Prompt: Aerial view rising over a wooden temple roof
surrounded by red maple forest, slow drone ascent, golden
hour light, light mist in the valleys.

Shot 5 - 5s
Prompt: Steam rises from a cup of matcha on a wooden veranda
overlooking a maple garden, close-up, slow push-in, soft
late-afternoon window light, warm film look.

Shot 3 is a 10-second walk, which is where longer-clip models earn their keep. Kling 4.0 handles longer durations, and Wan 3.0 is a lower-cost option for the scenic shots (1, 2 and 4). Generate shot 1 first. If you like its colour, ask the assistant to “match the colour description of shot 1 in every other shot” before you render the rest.

To skip the manual stitching, the AI storyboard generator and AI reel maker take a script or idea and build a multi-shot video in one flow. You can paste the assistant’s script straight into either.

The image-first workflow (better consistency)

Text-to-video invents the subject from scratch every time, so a character or product can look slightly different from shot to shot. The fix is to make a still first and animate it:

Take the “start-frame idea” line from each shot in the assistant’s output.
Generate that still in the AI image generator, attaching a reference photo (your product, your face, your pet) when it has to match something real. An image you made in ChatGPT works as a start frame too.
Open image-to-video, upload the still, and paste only the motion part of the shot prompt: the action and camera move. The look comes from the image, so you don’t need to describe it again.
Generate. Because every shot starts from an image you approved, the clips match far better.

This costs one image plus one video per shot. It also saves credits, because you stop regenerating videos whose first frame was wrong.

Which VdoBloom model for which shot

Shot type	Good first choice	Why
Product close-ups, premium ads	Kling 4.0	Clean detail and stable shapes, clips up to 30 seconds
People moving, dance, action	Seedance 2.5	Natural body motion and multi-reference control
Expressive characters, emotion	MiniMax H3	Strong facial performance and physical motion
Shots that need sound or dialogue	Veo 3.1	Generates audio with the video
Scenery, B-roll, testing ideas	Wan 3.0	Lower cost per clip, good for drafts and landscapes

Ask the assistant to tag each shot with one of these model names. Add this line to the director prompt: “Suggest a model for each shot from: Kling 4.0 (product detail), Seedance 2.5 (body motion), MiniMax H3 (faces and emotion), Veo 3.1 (needs audio), Wan 3.0 (scenery and drafts).”

Tips that save time and credits

Draft cheap, finish premium. Test each prompt on a lower-cost model or a shorter duration, then re-run the keeper on the model you want.
One change at a time. When a shot fails, tell the assistant what went wrong (“the bottle melted when the lid came off”) and ask for a simpler action, not a longer prompt.
Reverse-engineer what you like. Found a clip or image with the look you want? Run it through video-to-prompt or image-to-prompt, then paste the result into the chat and ask the assistant to write your shots in that style.
Keep text out of the frame. Add captions, prices and logos in an editor afterwards.
Use only people you have permission to show. Real faces in reference photos should be yours or belong to someone who agreed.

What it costs

VdoBloom works on credits with no subscription required. You buy a pack, the credits never expire, and every clip shows its credit cost before you press generate. Short drafts on lighter models cost a few credits, and longer, higher-resolution premium clips cost more. The assistant side costs nothing extra if you already use the free tier of ChatGPT or Claude. New accounts get free credits to try the workflow.

Frequently asked questions

Can ChatGPT make videos by itself?

Not in a normal chat. ChatGPT writes scripts, shot lists and prompts, and it can make still images. Turning those into moving video needs a video model, either through a third-party app or by pasting the prompts into a generator like VdoBloom.

Can Claude generate videos?

No. Claude writes text and reads images, but it doesn’t output video. It’s good at structured shot lists, so use the director prompt above and render the shots on a video platform.

Does VdoBloom have a ChatGPT app or Claude connector?

No. The workflow is copy and paste: the assistant writes the prompts and you paste them into VdoBloom’s text-to-video or image-to-video tool.

What’s the best prompt format for AI video?

One shot per prompt, 40–70 words, covering subject, one action, setting, camera shot size and move, and lighting. Repeat key visual details in every shot so separately generated clips match.

How do I keep the same character across shots?

Use the image-first workflow. Generate or upload one reference image of the character, make each shot’s start frame from it, then animate each frame with image-to-video.

How long can each clip be?

It depends on the model. Most clips run 5 to 10 seconds, and some models, like Kling 4.0, go longer. Plan your shot list in 5- and 10-second blocks and join the clips in an editor or with the storyboard tool.

Do I need a paid ChatGPT or Claude plan?

No. The free tiers handle the director prompt fine. Paid plans mainly give you more messages and longer conversations.

Do I need a subscription on VdoBloom?

No. You can buy a one-time credit pack, the credits never expire, and new accounts start with free credits.

Open the text-to-video tool, paste the first shot from your assistant, and check the credit cost on screen before you generate.