Make Animated Shorts from Scratch with Claude Code

Input a creative brief or reference video → Claude Code orchestrates Remotion/Codex/ElevenLabs to generate code-drawn animation → output MP4 with sound effects and voiceover

Who

Anyone who wants to make short animated ads or explainer videos without design or development background

You get

After environment setup in Claude Code, generate a code-based animation project renderable to MP4 via a detailed brief, reference video, or live-action footage

You need

Codex · ElevenLabs · FFmpeg

Steps

1. Check and install dependencies

   Open Claude Code and paste the environment check prompt. Claude will detect whether Node.js, Python (with numpy and opencv-python), ffmpeg, and the Codex CLI are installed, and return install commands for anything missing. If something is missing, ask Claude to help install it.

   Execution prompt · Source text。Claude Code chat input

   Copy prompt

   Check which of these are installed on my computer and their versions: Node.js, Python (with numpy and opencv-python), ffmpeg and the Codex CLI. Tell me what's missing and the install command for each one on my system.

2. Load rules and approval gates

   Before any creation, paste the rules and approval gates prompt into Claude Code. This prompt contains animation rules (readable without sound, fact-checking, code-drawn video, animation principles, budget tracking), 6 verification loops, and 5 mandatory human-approval checkpoints. Claude will pause at each checkpoint and show the work, not just describe it.

   Execution prompt · Source text。Claude Code chat input; paste before each creation request

   Copy prompt

   Rules:
   - The video must read with the sound off.
   - Only use facts I give you or that you can source. No made-up numbers or claims.
   - Use text sparingly and integrate it into the world (painted, stamped, written) rather than overlaying it.
   - Generated images are references only. They never appear on screen; draw the video in code.
   - When creating the storyboard, generate only 1 image first so I can check the style. Wait for my approval before generating the rest.
   - The timing comes from the audio. All animation timing reads from timing.json, with nothing hardcoded.
   - Anticipation before every big action, follow-through after, squash and stretch, springs for landings, arcs for every moving object, and no linear easing.
   - Held props and text never mirror when a character turns. Motivate every camera move.
   - Track every ElevenLabs credit and every image you generate, and never exceed [budget].
   
   Verification loops (do not skip any):
   1. Key-pose match: compare each key pose frame side by side with its key art and fix what differs.
   2. Sound-off read: a fresh subagent reviews a silent contact sheet at 2 fps and retells the story. If it can't, fix the staging and repeat.
   3. Sync check: every key action and voice line within 1 frame of its timing.json target.
   4. Fact check: every word on screen and in the voiceover matched to its source. Cut anything unmatched.
   5. Technical: exact duration, stereo audio, peaks below −1 dBFS, no unintended dead air.
   6. Director's review: watch it twice as frames, write a brutal critique, fix the top 3 issues, review again.
   
   Approval gates:
   1. Facts + story
   2. Style: 1 storyboard image first → wait for my approval. Then character sheets and key poses → wait for my approval
   3. Audio-only mix and timing.json
   4. Final video, contact sheet and critique
   5. Save the process as a skill
   
   Keep a task list throughout. At each gate, show me the work, not a description of it.

3. Create with a detailed brief (Way 1)

   Fill in the brief template placeholders: length, format (16:9 or 9:16), voiceover needs, story outline (ask for 2-3 options), character descriptions (shape, color, personality, movement), visual style (with hex codes), reference image/video, target audience, and ElevenLabs/image budget caps. Paste the completed brief into Claude Code, then append the rules prompt at the end.

   Execution prompt · Source text。Claude Code chat input; replace bracket placeholders before sending

   Copy prompt

   Make a [length]-second animated [ad / explainer / story] about [subject].
   Format: [16:9 or 9:16]. Voice: [narrator / no words].
   
   Story: [what changes from start to end]. Give me 2–3 options to pick from.
   
   Characters:
   - [Name]: [shape and colour], [one detail that shows who they are], [personality and how they move].
   - Mascot (optional): [what it is]
   
   Look: [style], colours [hex codes], [day/night, mood].
   Reference: [image/video], for the look only.
   Audience: [who's watching]. The first second must grab attention.
   Budget: max [X] ElevenLabs credits, max [X] images.

4. Create with a reference video (Way 2)

   Put the reference video in your project folder, then paste the reference video prompt, replacing length, topic, and video filename. Claude will analyze the reference frame by frame, extract style, pacing, and camera work, and build something new with it. For more control, append character descriptions, brand colors, story beats, or voiceover needs, then always append the rules prompt.

   Execution prompt · Source text。Claude Code chat input; replace bracket placeholders before sending

   Copy prompt

   Make a [length]-second animated video about [your topic] in the style of [video file]. Match its look, pacing and camera, but create a new story and new characters.

5. Reuse ready-made templates or skills (Way 3)

   Sub-path A — Ready-made prompts: Find a creator-shared animation prompt, swap in your topic, length, colors, and characters, paste into Claude Code, and append the rules prompt at the end. Read the prompt before running to confirm it uses the same tools (Remotion, ChatGPT, ElevenLabs) and remove anything unwanted. Sub-path B — Skills: Find a skill shared as a GitHub link or similar, ask Claude to install it: `Install this skill: [link]`, then invoke with a short request: `Use the [skill name] skill to make a [length]-second video about [subject]`.

   Execution prompt · Source text。Claude Code chat input; replace with actual skill link

   Copy prompt

   Install this skill: [link]

   Execution prompt · Source text。Claude Code chat input; use after skill installation

   Copy prompt

   Use the [skill name] skill to make a [length]-second video about [subject].

6. Turn live footage into animation (Way 4)

   Place your live-action video (product demo, screen recording, or owned footage) in the project folder, paste the live-to-animation prompt, replacing video filename, length, and style. Claude will split the video into frames, map the timeline, have Codex/ChatGPT redraw key moments in your chosen style, then animate to match the original. Append character conversion, brand colors, audio handling, or end text as needed, then append the rules prompt.

   Execution prompt · Source text。Claude Code chat input; replace bracket placeholders before sending

   Copy prompt

   Turn [video file] into a [length]-second animation. Keep the same scenes, movements and timing, and redraw everything as [style, e.g. 2D cartoon].

7. Save as a reusable skill

   After completing a video you're happy with, paste the save skill prompt to have Claude save the entire process as a reusable skill. Next time, make a new video with one short request.

   Execution prompt · Source text。Claude Code chat input

   Copy prompt

   Save everything we just did as a reusable skill, so next time I can make a new video with one short request.

Rules:
- The video must read with the sound off.
- Only use facts I give you or that you can source. No made-up numbers or claims.
- Use text sparingly and integrate it into the world (painted, stamped, written) rather than overlaying it.
- Generated images are references only. They never appear on screen; draw the video in code.
- When creating the storyboard, generate only 1 image first so I can check the style. Wait for my approval before generating the rest.
- The timing comes from the audio. All animation timing reads from timing.json, with nothing hardcoded.
- Anticipation before every big action, follow-through after, squash and stretch, springs for landings, arcs for every moving object, and no linear easing.
- Held props and text never mirror when a character turns. Motivate every camera move.
- Track every ElevenLabs credit and every image you generate, and never exceed [budget].

Verification loops (do not skip any):
1. Key-pose match: compare each key pose frame side by side with its key art and fix what differs.
2. Sound-off read: a fresh subagent reviews a silent contact sheet at 2 fps and retells the story. If it can't, fix the staging and repeat.
3. Sync check: every key action and voice line within 1 frame of its timing.json target.
4. Fact check: every word on screen and in the voiceover matched to its source. Cut anything unmatched.
5. Technical: exact duration, stereo audio, peaks below −1 dBFS, no unintended dead air.
6. Director's review: watch it twice as frames, write a brutal critique, fix the top 3 issues, review again.

Approval gates:
1. Facts + story
2. Style: 1 storyboard image first → wait for my approval. Then character sheets and key poses → wait for my approval
3. Audio-only mix and timing.json
4. Final video, contact sheet and critique
5. Save the process as a skill

Keep a task list throughout. At each gate, show me the work, not a description of it.