Skip to content

A Claude Code plugin that turns a story into a finished anime

Ruby
README

Make an Anime

A Claude Code plugin that turns a story into a finished anime: character sheets, camera-directed shots, voices, a score, titles, subtitles, sound effects and finishing VFX. It works at any length, from a 12-second scene to a full episode.

Four frames, one from each example: two friends by a vending machine, a scientist and his assistant in a lab, a Muay Thai fighter against a karate fighter at sunset, and a grandmother and granddaughter in an omelette shop

Examples

All four were made with the plugin, with Japanese voices and English subtitles, and they show how far the same steps can stretch. Sound on.

The Last One: slice of life, 12 s

Two university friends have just missed the last train home. Across the rice fields it shrinks into the distance, and the only question left is what to get from the vending machine.

https://github.com/user-attachments/assets/4be1633c-94f1-4c79-97c5-e8a69d96edc2

The Invention of the Century: comedy, 20 s

A theatrical scientist unveils a shrinking fluid in his cramped lab. His deadpan assistant thinks it's melon soda, and then she eats the evidence.

https://github.com/user-attachments/assets/7ec2cd36-a097-406c-b1a9-072a2deb27ed

One Round at the River: fight, 17 s

A cocky Muay Thai fighter challenges a calm karate man to one fight on a riverbank at sunset, with hard hand-to-hand contact and no superpowers.

https://github.com/user-attachments/assets/9be8bfb9-86dc-4d77-ba8d-c2e7a2b92733

Too Salty: cooking drama, 74 s

For 91 days, Hana has tried to make her grandmother's rolled omelette, and every morning she hears the same verdict. Then one morning she leaves out the salt. This is the longest of the four, a whole short story told in 19 shots, painted with GPT Image 2.6.

https://github.com/user-attachments/assets/352d902c-39b4-4511-8496-56b03737c028

The toolkit also works at full length: it was first used for a narrated anime documentary of over 12 minutes, in about 40 scenes. See Full episodes.

Installation

In Claude Code:

/plugin marketplace add makevoid/claude-make-an-anime
/plugin install make-an-anime@make-an-anime

The plugin asks for your Fal API key and stores it as a sensitive option. All paid generation goes through Fal, so you need an account with credit.

You also need these on your machine:

Tool Used for
Ruby 3.2+ and Bundler The toolkit itself
FFmpeg, ImageMagick Video, audio and image work
Python 3 Green-screen keying, score editing, analysis
Node.js 22.12+ and Chrome Rendering p5 titles, subtitles and composites
macOS 14+ and Swift 5.9+ Optional: the Core Image VFX pass

Each project installs its own Ruby, Node and Python packages when the skill runs setup, and doctor reports anything missing.

Usage

Ask Claude for what you want:

/make-an-anime Make a 2-minute pilot from story.md, with characters based on the photos in ./refs

Claude reads the story, writes a production plan (cast, voices, shot list, every prompt, the number of paid calls) and asks you to approve it. No paid call runs before that approval. It then produces and reviews each step, and retakes shots that fail review.

Reference photos are optional: original characters are designed from text. Use photos of real people only with their consent.

MiniMax H3 remains the default video model. Ask explicitly for Seedance to use Seedance 2.5 US, defaulting to 480p unless you specify another resolution. It supports image-to-video, reference-to-video and optional draft completion at 1080p. See Seedance settings and drafts.

How a scene is made

Each stage reads prompts from prompts/<run>/ and writes to output/<run>/, so everything can be inspected and redone on its own.

One shot of Too Salty at four stages: the character sheet, the painted keyframe, a frame of the animated take, and the finished frame with subtitles
  1. Character and location sheets. A style bible (config/style.md) guides the image prompts, so the whole cast shares one look. GPT Image 2.6 draws a sheet for each character and location, and those sheets become the references for everything after.
  2. Voices and score. Each character gets one ElevenLabs voice in config/cast.yml. Voices come before animation because line lengths decide shot lengths. ElevenLabs Music then composes one score, and gen:mix ducks it under the voices.
  3. Keyframes. GPT Image 2.6 paints the first frame of each shot from the sheets.
  4. Shots. MiniMax H3 animates each keyframe as one continuous take, with the camera move written into the prompt and the speaker's voice as target audio so the mouth follows the line. Held stills and colour inserts cost nothing. Existing takes can be reused across scenes with from_take. For moments that need exact placement, layered scenes composite green-screen sprites over a painted plate.
  5. Overlay. A p5 script (05_overlay.js) draws titles, subtitles, fades and any composite work.
  6. Sound. sfx:gen creates sound effects with ElevenLabs Sound Effects v3, and sfx:mix places them from a cue list and levels them against the mix.
  7. Finish. vfx:render adds grain, glow or impact effects with Swift Core Image, on frame-accurate cues. The optional vfx:bake pass preserves the plain picture so repeating the finish does not accumulate effects.

For detailed staging and playback review, see Staging and review. Local graphics tools include deterministic particles and light-only glow, plus animation asset preparation.

The same steps as commands, once a project and its prompts exist:

export RUN=my-scene
ruby mv.rb gen:dialogue
ruby mv.rb gen:music
ruby mv.rb gen:mix
ruby mv.rb gen:keyframes
ruby mv.rb gen:shots
ruby mv.rb gen:overlay
SFX=my-scene-sfx ruby mv.rb sfx:gen
SFX=my-scene-sfx ruby mv.rb sfx:mix
ruby mv.rb vfx:build
VFX=my-scene-vfx ruby mv.rb vfx:render    # -> output/my-scene/final_vfx.mp4

Check current prices on Fal before you run anything. For scale, the first build of Too Salty, made with GPT Image 2.5, took 4 sheets, 22 keyframe images, 23 H3 takes, about 22 voice takes, 1 score and 9 sound effects, retakes included.

How a scene is edited

Everything is a file and every step is cached, so a change only redoes what it affects. Replaced takes are kept in output/<run>/rejected/. These are real edits from the examples:

What was wrong The edit Cost
Two voice lines overlapped Changed at: in 03_dialogue.yml, then gen:dialogue and gen:mix reused the takes Free
A voice take said its word twice Added gate_s to mute the first one Free
A tear ran down Hana's cheek Repainted the keyframe and changed the prompt, then ONLY=outburst FORCE=1 on gen:keyframes and gen:shots 1 image and 1 H3 take per try
An umbrella vanished from the background mid-shot Painted an end frame that keeps it, and set end_image on the shot 1 image and 1 H3 take per try
Steam washed out a tray of omelettes Replaced the take with a slowly zooming still of its keyframe Free

A retake needs a changed prompt. Fal requests are stored as receipts by their exact input, so FORCE=1 with an unchanged prompt fetches the same result again, at no charge.

Local review tools help decide what to retake:

Command What it shows
review:camera The camera move each shot made, with frame strips to read by eye
review:dialogue What Whisper heard in each take (paid)
dialogue:timeline Lines against shots and cuts
audio:seams Abrupt music changes at scene joins and inside a score

Full episodes

A longer project uses the same steps, one run per scene, plus:

  • scene:new[ep01-s05,480] to scaffold a scene, with one config file per scene so several agents can work at once
  • config/episodes/ep01.yml for the playing order
  • episode:preview to join scenes, levelling the voices across them (VOICE_DB) and fading the seams (SEAMS=1)
  • production waves with sub-agents (work:next, work:accept), starting with the opening and the emotional peak

full-episode.md has the complete workflow.

Repository layout

.claude-plugin/                   plugin and marketplace manifests
skills/make-an-anime/
  SKILL.md                        what Claude follows
  references/                     planning, prompts, camera language, audio, episodes
  examples/                       workflow showcases only: 01 is read on every run; 02-04 are three of the films above
  scripts/
    mv.rb, mcp.rb                 CLI and MCP entry points
    lib/fal/                      Fal queue client with receipts, and model wrappers
    lib/pipeline/steps/           RefBase, Dialogue, Music, Mix, Keyframes, Shots, Clips, Overlay
    lib/media/, lib/toolkit/      FFmpeg, ImageMagick, Python, p5 and Swift wrappers; project tools
    tools/p5/                     headless renderer and the Anim library
    tools/python/                 keying, tracking, analysis, score editing
    tools/vfx/                    Swift Core Image renderer
    spec/                         RSpec suite

The specs never call Fal unless you authorize the live spec. They run from skills/make-an-anime/scripts after npm ci, with MV_PYTHON pointing at a Python that has requirements.txt installed:

bundle exec rspec --tag core
bundle exec rspec --tag media    # includes a whole scene with synthetic media

Bonus: behind the scenes

Everything here comes from the four examples above, plus one extra film at the end: what the skill generated, what got thrown away, and what changed on the way to the finished film. The demo videos show the first build of Too Salty, made with GPT Image 2.5. Its sheets, keyframes and frame strips further down come from the current GPT Image 2.6 version.

Demo videos

Raw MiniMax H3 takes. 18 takes straight from the model, with no subtitles, grading or sound.

https://github.com/user-attachments/assets/49479813-9511-44c2-94a1-74e501aab90e

One shot, three stages. One shot from each anime, first as the painted keyframe, then as the silent H3 take, then finished with sound on.

https://github.com/user-attachments/assets/11ecdaf9-3d56-4855-bff2-ddfa5c6a423b

Rejected takes next to the kept ones. Seven real retakes and what fixed each one. Silent.

https://github.com/user-attachments/assets/3b14dde3-efe6-40db-9787-b4cd3c1998c6

The same scene in two languages. The Last One was made in English first, then voiced again in Japanese with English subtitles.

https://github.com/user-attachments/assets/638400e6-5f41-445d-b2cd-281a6d3dcd3b

Character sheets

The image model draws a turnaround and expressions for each character. Every later image and take uses these sheets as references, which keeps the characters on-model from shot to shot.

The Last One: Mio and Haru

Character sheet of Mio: glasses, yellow shirt, green skirt and backpack Character sheet of Haru

The Invention of the Century: the scientist and Aoi

Character sheet of the scientist: lab coat, red shirt, goggles on his head Character sheet of Aoi: blue hair, glasses, lab coat over a hoodie

One Round at the River: Riku and Ren

Character sheet of Riku: black tank top, red Muay Thai shorts, hand wraps Character sheet of Ren: white karate gi with a black belt

Too Salty: Hana and Kiku

Character sheet of Hana: headscarf, pale shirt, dark apron Character sheet of Kiku: elderly woman in a dark top and white apron

A character is cheapest to change at this stage. The scientist went through three sheets, and the shots made from an earlier one were redone from the last:

Three versions of the scientist's character sheet, from scruffy and grimy to clean-cut

Location sheets

Locations get sheets too, so a place stays the same from one shot to the next.

A country road through rice fields below green mountains A cramped laboratory at night, full of shelves and glassware

A grassy riverbank and an iron bridge at sunset A small omelette shop at night, with a counter and a copper pan

A narrow old shopping street, wet from the rain

Keyframes

Each shot starts from one painted frame. These are the keyframes of all four anime.

The Last One

Three keyframes of The Last One

The Invention of the Century

Eight keyframes of The Invention of the Century

One Round at the River

Nine keyframes of One Round at the River

Too Salty

Sixteen keyframes of Too Salty

MiniMax H3 takes, frame by frame

Five frames across each continuous take, in order.

Two characters act at once: he keeps talking while she eats the evidence.

Five frames of the scientist talking while Aoi eats a tiny pineapple with tweezers

A start frame and an end frame are both painted, and H3 animates the change between them.

Five frames of a pineapple shrinking to almost nothing under a dropper

The camera moves around the fighters as they trade blows.

Five frames of the two fighters exchanging punches while the camera circles them

The lips follow a whispered line, and the two hands stay on the pan.

Five frames of Kiku and Hana with their eyes closed over a pan, Kiku whispering

Cooking is the hard part to animate: the omelette has to roll as one piece.

Five frames of a rolled omelette being pushed along a copper pan

A wide shot with a small action: the shop curtain goes up at dawn.

Five frames of Hana hanging the curtain over the shop door on a wet street

Editing takes together

The first two frames come from separate H3 takes, and the last is the finished split panel, cut together in p5 with the subtitle on top.

A frame of Riku, a frame of Ren, and the finished split panel of the two with a subtitle

One more film: Final Inspection

A rusted 40-metre robot is due to be cut up tomorrow. Tetsu, the old mechanic who kept it running for 30 years, climbs up for one last check and shares a thermos of tea with it. 34 s, painted with GPT Image 2.6.

https://github.com/user-attachments/assets/b91a89cc-f32b-46f3-95ac-fcf6d19f8e58

License

MIT