AI animation in Make: Claude turns a character and an episode idea into a 30-second animated story (free scenario template)

Hey Make Community, AI video makes animated stories easy to start, but hard to turn into a series: draw the same character twice and you get two different characters. So I built a free Make scenario that keeps them the same: list your episodes in a CSV in Google Drive, and for each one you get a 30-second animated vertical episode with narration, captions, a title and music. Your character is drawn once, and every scene of every episode is drawn from that one picture, so they look the same from episode to episode. A Claude agent writes the story and designs each episode, Shotstack draws, animates and renders it, and the MP4 lands in your Google Drive.

The blueprint, ready to import:

episodes-blueprint.json (46.1 KB)

How the scenario works

  1. Download the CSV reads your episodes file from Google Drive: the character, a link to a picture of them if you have one, what happens in the episode, and the style.
  2. Read and check the rows is a Make Code step. It turns the CSV into rows and stops the run before anything costs money if a character is a child (the video models refuse to animate children) or a picture isn’t a link. It also lists each character once.
  3. One drawing per character and Draw the character use Shotstack’s Generate an Asset module to draw one clean reference picture per character with GPT Image 2.5 Sunburst, from your description, or from your own picture with GPT Image 2.5 Sunburst Edit. A wait loop checks until each drawing is done.

  1. The drawings (a Text aggregator) collects the pictures, and Add the drawings gives every episode the picture of its character. Rows with the same character share one drawing, so a series stays consistent.

  1. The brief writes the instructions for each episode in plain English: the story shape, the narration, the 5 scenes drawn from the reference, the animation, the captions and the music. It’s a Make Code step, so the brief is easy to read and change.

  1. Claude makes the video starts a Claude agent session (Claude Opus 5.5 at extra-high effort, through Anthropic’s Managed Agents). The agent works in its own cloud sandbox with Shotstack’s command-line tool and its agent skill installed. It writes the story, builds the whole episode as one Shotstack edit, then runs a check script and Shotstack’s validator on it and fixes the edit until both pass.

  1. Make looks in on the session once a minute (a Repeater, a Sleep and a status check), then reads the finished edit from the session’s files. Check the edit runs the same check once more before anything is rendered: every scene picture is drawn from the reference, with the right model settings.

  1. Make the video hands the edit to Shotstack’s Render from JSON module. In that one render, Shotstack draws the 5 scene pictures from the reference with GPT Image 2.5 Sunburst Edit, animates each one with Wan 3.0 Prime, narrates the story with ElevenLabs, and adds the title, the captions and the music.

  1. Make checks the render once a minute, downloads the MP4 and saves it to My Drive, named after the episode idea. If the edit fails the check, or Shotstack says the render failed, a router writes a - failed.txt note to Drive with the reason instead.

The first run also sets itself up: if your Anthropic account doesn’t have the agent yet, the first route creates it (“Shotstack video editor”) and its sandbox (“Shotstack CLI”), and every run after that reuses them.

Claude’s sandbox has no Shotstack key, so the agent can only write and check the edit. The drawings, the render, and what they cost, stay in Make on your own Shotstack connection.

Here’s the brief Claude got for the first sample episode, starring Pip the robot:

The brief
Make a 30-second vertical story episode starring this character: an animated short people watch to the end on TikTok, Reels and YouTube Shorts.

The character: Pip, a small round robot with one big blue eye and wheels for feet, who collects shiny things
Their reference picture, how they always look: (the picture of the character that Draw the character made)
This episode: Pip finds a lost key in the park and follows the clues to its owner
Style: 3D animated film

The story
- Write it first: a first line that hooks, then a small problem, a turn, and a warm or funny ending that pays off the hook, told in 5 scenes. Plain, vivid words in the language of the episode idea.
- Narration: one audio clip generated with elevenlabs-multilingual-v2, alias "voice", start 0.5, length "auto", 60 to 70 words, read by a storyteller voice from the guide that suits the story. Numbers spelled out.

The pictures
- 5 scene images, one per scene, made with gpt-image-2.5-sunburst-edit, with options imageUrls holding only the reference picture URL above, aspectRatio "9:16" and resolution "1K", so the character looks exactly the same in every scene.
- This model edits from the reference, so each prompt first says what to keep: the character's face, body, colours and clothes exactly as in the reference picture. Then it says what to change: the action, the place, the light and the camera framing, and how anything the story depends on looks at that moment (open or shut, found or lost, full or empty), in the 3D animated film style, with no words in the picture.
- Give the image clips the aliases "scene1" to "scene5" and put them on the lowest track, each at the same start and length as its video, so they are never seen.
- Bring each scene to life with wan-3.0-prime-image-to-video: options startSrc "alias://scene1" and so on, aspectRatio "9:16", resolution "720p" and generateAudio false, volume 0, and a prompt for the motion and the camera move. Lengths of 5 to 7 seconds, back to back from 0 on one track. Count about 2.5 words a second, and make the videos run about 2 seconds longer than the narration.
- Anyone in a picture is an adult.

Words on screen
- The episode title for the first 2.5 seconds: rich-text, 6 words or fewer, in Anton, centred in the upper third.
- Captions: a rich-caption clip with src "alias://voice", start and length "alias://voice", width 960, height 200, fit "none", in the lower third: two or three words at a time, Montserrat ExtraBold at weight 800 with style letterSpacing 4, white with a dark stroke, the spoken word in one accent colour taken from the style. Keep other words out of that band.
- Load the fonts in timeline.fonts and use the file name as font.family: Anton https://fonts.gstatic.com/s/anton/v27/1Ptgg87LROyAm0K08i4gS7lu.ttf and Montserrat ExtraBold https://fonts.gstatic.com/s/montserrat/v31/JTUHjIg1_i6t8kCHKm4532VJOt5-QNFgpCvr70w-.ttf
- In the last 2 seconds, a short line inviting people to follow for the next episode, in the language of the story.

Sound
- Music: one audio clip from Shotstack's free music library, start 0, length "end", volume 0.12, under the narration. Pick the track whose name best fits the mood: https://shotstack-assets.s3.amazonaws.com/music/unminus/lit.mp3, https://shotstack-assets.s3.amazonaws.com/music/unminus/happy.mp3, https://shotstack-assets.s3.amazonaws.com/music/unminus/ambition.mp3, https://shotstack-assets.s3.amazonaws.com/music/unminus/palmtrees.mp3, https://shotstack-assets.s3.amazonaws.com/music/unminus/berlin.mp3. Do not generate music.

Easy to miss
- An alias works only as a startSrc or endSrc, never inside imageUrls: imageUrls hold the reference picture URL.
- tracks[0] is the top layer.
- Only clips that start at 0 may use the length "end".
- Output: mp4, 1080 x 1920, 30 fps.

Reply with the edit JSON only: no code fence, no comments.

What’s in each episode

  • The same character in every scene. Each scene picture is drawn from the one reference picture: the prompt says what to keep (the face, the colours, the clothes) and what changes (the action, the place, the light).
  • A real little story. A hook, a small problem, a turn and a warm or funny ending, told in 5 scenes and narrated by a storyteller voice.
  • Five animated scenes, each brought to life by Wan 3.0 Prime with its own motion and camera move.
  • A title, word-by-word captions and a closing line inviting people to follow for the next episode.
  • Any style: 3D animated film, storybook watercolour, anime, claymation or paper cut-out.

No fixed layout: Claude writes and designs every episode

There’s no video template. Claude follows the brief’s rules and Shotstack’s agent skill, and writes the story and every scene itself. Run the same idea again and you get a different episode.

How to set it up

It takes about 10 minutes. You need a paid Make plan: an episode takes 15 to 25 minutes, and Make’s Free plan stops a run at 5 (Core allows 40).

  1. Get a Shotstack API key, and an Anthropic API key with access to Managed Agents (beta) from platform.claude.com.
  2. In Make, create a new scenario, click ⋯ (More) at the bottom of the editor, choose Import Blueprint and pick episodes-blueprint.json.
  3. Add your connections: Google Drive on the Drive steps, Anthropic Claude on every Claude step, and Shotstack on the Shotstack steps.
  4. Make a CSV with the columns character, picture, idea and style. Describe the character, add a link to a picture of them if you have one (or leave it empty and AI draws them), write what happens in the episode, and pick a style. Upload it to Google Drive and pick it in Download the CSV. The sample (two episodes with the same character):
character,picture,idea,style
"Pip, a small round robot with one big blue eye and wheels for feet, who collects shiny things",,Pip finds a lost key in the park and follows the clues to its owner,3D animated film
"Pip, a small round robot with one big blue eye and wheels for feet, who collects shiny things",,"Pip's shiny collection rolls away down the hill on a windy day, and Pip has to win it back before the rain",3D animated film
  1. Click Run once. The episodes land in My Drive.

What it costs and how long it takes

What Cost or time
Shotstack About 30 credits an episode: 5 animated scenes, 5 scene pictures, the narration and the render, plus one drawing per character
Claude Billed by Anthropic for the agent’s session. The brief step caps each session at $20, and you can change it
Make My test run used 192 operations (198 Make credits)
Time My test run took 25 minutes: Claude takes about 10 to write and check the edit, and Shotstack several minutes to draw, animate and render

Ideas

  1. Read your episodes from Google Sheets and schedule the scenario, so a new episode comes out every week.
  2. Swap the animation model in the brief: MiniMax H3 costs less, Seedance 2.0 and 2.5 move more smoothly.
  3. Add your own picture link to each row to turn your mascot or a product character into a series.

Good to know

  • Start with one episode per run: an episode takes 15 to 25 minutes, and Make stops a run at 40 minutes on Core.
  • Video models refuse to animate children, so make the character an adult, an animal, a creature, a toy or a robot.
  • Each character is drawn once per run and reused in all its episodes in that run.
  • The scenes are AI-generated: label the episodes as AI-generated when you post them.
  • Tested on Make with Claude Opus 5.5 at extra-high effort.

If you try it, post your episode in the thread, I’d love to meet your character. And what should I build next?