How to make AI videos with consistent characters in Claude
Short answer: AI video models invent a slightly different face every time you ask for a new clip. To keep one character across a whole story, do not go straight from text to video. Make a character sheet first, then one still image per shot using the sheet as a reference, check the face in every still, redo the ones that drifted, and only then turn each still into a short video. In Claude you can do all of it in one chat with a video connector. This guide uses Dora's, because it is ours. It is the same storyboard method (character sheet, panels, face check) we use for our own Bible videos.
Dora is ours. The steps below are the method; the example prompts are written for this guide and were not run for it.
What you need
- Claude with a video connector. Add Dora's under Customize > Connectors with the address https://doravideo.com/mcp; the full steps are in how to make videos in Claude.
- A Dora plan. See pricing.
- A short story in shots. Three to six shots is a good first project. Write one line per shot: where, what the character does, how close the camera is.
Why faces drift
Each generation starts from noise. If the only thing tying two clips together is a description ("a woman with curly dark hair"), the model fills in everything else differently each time: face shape, age, clothes. A reference image fixes that. Dora's image models Nano Banana 2 and Nano Banana Pro, and the Seedance 2 and Seedance 2 Fast video models, accept up to 7 reference images, most important first; the tool descriptions tell Claude to use them for "character face/body anchors".
Step 1: the character sheet
Ask for one image that shows the character from several angles on a plain background. Describe things that stay the same in every shot: hair, clothes, one or two details.
"Make a 9:16 image: character sheet of Maya, a woman in her early 30s with curly dark hair tied up, a terracotta linen apron over a cream t-shirt and small silver hoop earrings. Front, side and three-quarter views on a plain light background, natural light, photorealistic. No text, no signs, no logos, no labels."
For faces, ask Claude to use Nano Banana Pro, the image model Dora lists as best at keeping a face from a reference photo. Keep the sheet's link: every later step uses it.
Step 2: one panel per shot
Now ask for one still image per shot, each time passing the sheet as the reference. Put the face in every panel, at different distances and angles, so you can check it.
"Using the character sheet as the reference, make a 9:16 image: medium shot from inside a small corner coffee shop at sunrise, Maya unlocks and opens the plain glass door, her face clearly visible, warm light. No text, no signs, no logos, no labels."
"Same reference: close-up of Maya steaming milk in a metal jug at the espresso machine, three-quarter view of her face, soft morning light. No text, no signs, no logos, no labels."
"Same reference: Maya behind the counter handing a plain white takeaway cup to a customer and smiling, medium shot, background customers soft and out of focus. No text, no signs, no logos, no labels."
Why "no text": image models write fake letters on glass, cups, jars and books. Saying so in every prompt saves redoing panels.
Step 3: the face check
Before any video, put the sheet and the panels side by side and look at three things: face shape, hair, and the one or two details you chose (here, the apron colour and the earrings). Ask Claude to compare them too, but decide yourself.
If a panel drifted, redo only that panel with the same prompt and the sheet as the reference, or add the best previous panel as a second reference. Fix faces in the stills, before any video: a clip that starts from a wrong face will not correct it.
Be honest with yourself about the hit rate. Some panels come out right the first time and some need two or three tries. Keep the rejects: comparing them teaches you which details to add to the prompt.
Step 4: one video per panel
Turn each approved panel into a short clip, with the panel as the first frame and one simple movement per clip.
"Make a 9:16 video from panel 1: slow push-in as she opens the door and steps inside, morning light flares on the glass. No text, no signs, no logos, no labels."
"From panel 2: she steams the milk, glances up at the camera and smiles, gentle steam, camera still."
"From panel 3: she hands over the cup and nods, the background stays soft and out of focus."
Dora's default video model is Seedance 2 Mini: vertical, 8 seconds, 480p, with sound. Ask for 720p if the clip is for a feed, and for "ambient sound, no dialogue" if nobody speaks. Claude gets a link to each file when it is done.
Watch each clip frame by frame, especially where the head turns: faces can still change mid-clip. If it changes, redo the clip from the same panel.
Step 5: put the clips together
The connector makes clips; it does not edit. In the Claude desktop app with Claude Code, you can ask Claude to download the clips and join them in order with ffmpeg. In the chat, download them and join them in any phone editor. See faceless Reels and TikToks with AI for captions and voice.
Mistakes that break consistency
- Changing the outfit mid-story without a new reference. If the character changes clothes, make a second sheet.
- Two new things in one shot. A new place and a new action and a new angle at once is where faces drift most. Change one or two.
- Tiny faces. A face that fills a few pixels cannot be checked and the model will not keep it. Keep the face visible in every panel.
- Skipping the check. Check the stills before any video, then check the clips too. A video costs more than an image, so fix faces in the stills first.
Frequently asked questions
Can Claude keep the same character across several AI videos?
Claude has no video model, but with a video connector it can pass reference images to the video service. Make a character sheet, use it as the reference for one still per shot, check the faces, then animate each still. That keeps the character much closer than text alone.
How many reference images can I use?
Dora's Nano Banana 2 and Pro image models and its Seedance 2 video models accept up to 7 reference images, most important first. The character sheet goes first; a previous good panel can go second.
Does it work the first time?
Not always. Some panels need to be redone. That is why the face check happens on still images, before any video is made.
Does the connector edit the clips together?
No. It makes the clips and gives you links. Join them in an editor, or ask Claude Code to do it with ffmpeg.
Sources
- Get started with custom connectors using remote MCP, Claude Help Center, checked November 17, 2026
- Dora MCP server specification, Dora, November 17, 2026