Lemo-Opuscar is a new open-source Claude Code skill that produces short films with no video-generation model at all. Instead, the agent writes rendering code: it draws every frame on a canvas or in WebGL, builds a score from free sample libraries, adds text-to-speech narration and stitches it together with ffmpeg. The repository, published by the developer Lemomo as lemomo-ai/lemo-opuscar, had about 2,000 stars and 211 forks when we checked on October 12, 2026, having been created on September 26.
The pitch is simple: pick one of 43 styles, from Chinese ink wash to 16-bit pixel RPG, and ask Claude Code for a film. The README says every demo was "directed, drawn, scored, and mixed" by an AI agent writing code, using Claude Opus 5.5. This post explains how it works, how to try it, what it realistically costs, and where the approach stops being impressive.
TL;DR: the questions people ask first
| Question | Short answer |
|---|---|
| What is it? | A Claude Code skill and plugin with 43 film styles, each a style prompt plus a demo film |
| Does it use a video model? | No. Canvas, WebGL, ffmpeg, sample-library music and text-to-speech |
| License? | MIT for the repo; third-party demo assets keep their own licenses |
| What do I need? | Node 20+, ffmpeg, Python 3.11+ (or uv); the agent installs the rest |
| How long per film? | Roughly 30 to 60 minutes of agent time, plus a lot of tokens |
| Default output? | 1920 by 1080 at 24 fps, other sizes on request |
| Which model? | Tuned for Claude Opus 5.5; others may not reproduce the styles |
| Best-known demo? | OPUSCAR 98, a 6:31 film redrawing all 98 Best Picture winners |
What does "a film made without a video model" actually mean?
Most AI video tools sample pixels from a diffusion or autoregressive model. Lemo-Opuscar does the opposite. The agent acts as director and programmer: it plans shots, writes JavaScript that renders each frame, runs it, checks the result, and iterates. The look comes from the code, not from a model that has seen millions of clips.
Filmstrip with one frame turning green, representing frame-by-frame rendering in a code-made film
That has real consequences. Output is deterministic and editable: if a title card is wrong, the agent changes a line of code rather than re-rolling a generation. Style consistency across shots, a notorious weakness of generative video, comes almost for free because one script defines the palette, line weight and camera rules. The cost is that the visual vocabulary is limited to what code can draw: flat shapes, procedural textures, typography, simple 3D. Photoreal footage is out of reach.
The README lists the techniques as frame-by-frame canvas and WebGL rendering, original music assembled from free sample libraries, and text-to-speech narration. For readers who have followed SWE-Game, the benchmark that asks coding agents to build Godot games, this is a neighboring idea: treat creative output as a software engineering task and let the agent's coding ability carry it.
The 43 styles
The styles are grouped into nine families in the README:
| Family | Count | Examples |
|---|---|---|
| Hand-drawn and painting | 7 | Crayon Picture Book, Watercolor Brush, Chinese Ink Wash, Impasto Oil Painting, Whiteboard Explainer |
| East Asian traditions | 4 | Shadow Puppetry, Ukiyo-e, Red Paper-cut, Paper-cut Lightbox |
| Print and printmaking | 5 | Risograph, Halftone Dossier, Woodcut, Copperplate Engraving, Silkscreen Travel Poster |
| Graphic and type | 7 | Swiss Motion Graphics, 60s Spy Title Sequence, Art Deco, Blueprint, Stained Glass |
| Information and keynote | 5 | Data Storytelling, Isometric Infographic, Dark Tech Keynote, Living Screencast, Sci-fi Hologram HUD |
| Cartoon and anime | 4 | 1930s Rubber Hose, 80s Cel Anime, Sci-Fi Sitcom Toon, Mid-century Cartoon |
| Games | 4 | 16-bit Pixel RPG, HD-2D, Microgame Frenzy, Game Show Flat |
| Cinema and eras | 2 | 1920s Silent Film, Liminal Found Footage |
| Materials and 3D | 5 | Brick Toy, Paper Pop-up Book, Tilt-Shift Miniature, Low-poly Isometric Island, Glass Product Render |
Each style lives in its own styles/<style>/STYLE.md. That file is the style prompt: it tells the agent how lines, color, motion and sound should behave. Two shared guides sit above them. DIRECTOR.md covers storytelling, sound, rhythm, camera, performance and self-checks, while TECHNIQUE.md covers rendering, voice, music and mixing. The agent reads the relevant ones before it starts.
The flagship demo: OPUSCAR 98
The headline demo is a 6:31 film that redraws all 98 Best Picture winners from 1927 to 2025, each in a style chosen to suit that film. The README breaks the set down as 77 films in 2D, 4 in 2.5D and 17 in 3D, and publishes the data as JSON along with a 1080p download. Other demos include a one-take screencast of the Claude Code mascot moving into the Claude app, a moon-base AI that answers Earth by drawing the word "home", a Microgame-style boot camp that speeds up until a final landing, and a product render of fictional glass earbuds.
These demos work as both showcase and test suite: each proves a style prompt can carry a full film, and each ships with the shot-by-shot direction so you can see how the agent reasoned.
How do I try it?
There are two routes, both from the README.
Plugin (recommended):
claude plugin marketplace add lemomo-ai/lemo-opuscar
claude plugin install lemo-opuscar@lemolab
On first use, the guides, tools and style prompts download to ~/lemo-opuscar, shared across films. Each film's project goes in the folder where you started Claude Code. Other agents can use the plugin/skills/lemo-opuscar/ folder by copying it into their own skills directory.
Clone: clone the repository and run claude inside it; films are written to films/<name>/.
Then just ask. The README's example request is a 45-second watercolor film about a family coffee farm with a warm female narrator. The skill asks follow-up questions about any unspecified topic details, whether you have your own voice, music or other material, and whether you want a storyboard first. If you do, it pauses once to show key shots.
Requirements: Node 20 or newer, ffmpeg, and Python 3.11 or newer (or uv). The agent installs the remaining dependencies. To update, run the plugin marketplace and plugin update commands and restart Claude Code; to remove it, uninstall the plugin and delete ~/lemo-opuscar.
Library shelf of cards with one pulled out, standing for the Lemo-Opuscar style library of 43 film styles
What does it cost?
The software is free, but the compute is not. The README is upfront that a film takes roughly 30 to 60 minutes of agent time and "a considerable number of tokens." There is no published token figure, so budget accordingly if you pay per token rather than on a flat plan. A short 45-second piece is a sensible first test; the 6:31 OPUSCAR 98 film is a different scale of effort. If you are already watching spend, our guides on Claude Code effort modes and token-saving skills are relevant, since a long agent session like this is exactly where context and effort settings matter.
How is this different from other Claude Code skills?
Most skills we have covered change how an agent codes or communicates. Show Me pushes visual output in coding sessions, and Anatomy generates isometric interactive figures. Lemo-Opuscar is closer to Anatomy in spirit, a skill whose product is a finished media artifact, but its scope is a whole timeline with sound. It is also a good example of the skill format being used as a distribution mechanism for taste: the 43 style files are essentially encoded art direction.
It also sits in a different place than video models. Use a generative model when you need photoreal footage or natural motion. Use something like this when you need consistency, editability, explainer-style visuals, motion graphics or stylized shorts, and when you want the source code of the film in your repository.
Limitations and honest caveats
- Model dependence. The README says the styles are tuned for Opus 5.5 and other models may not reproduce them. Treat results on other agents as experimental.
- Setup friction. You need several local tools, and ffmpeg and Python environments are a common place for things to go wrong. The README asks stuck users to open an issue.
- Time and tokens. Half an hour to an hour per film is not interactive. Iterating on a storyboard is cheap; re-rendering a full film is not.
- Asset licensing. The MIT license covers the repo. Third-party assets in the demos keep their own licenses, listed in per-demo CREDITS files, and you are responsible for anything you bring in.
- Quality ceiling. Code-drawn films look like code-drawn films. That is the point for many styles, and a limit for others.
- Young project. About 24 commits on main at the time of writing. Expect the structure to move, and check the repository for current issue counts.
What this signals
Two trends meet here. First, agent skills are growing from small prompt snippets into large packaged libraries with guides, tools and examples, installable through a plugin marketplace. Second, coding-capable models are being used as general production engines: if a model can write a renderer, it can make a film, a slide deck or an interactive figure. Our guide to agent skills explains the format, and the Claude Code commands reference covers the plugin commands used above.
If you build with skills, the takeaway is practical: a skill's value is often in its examples and rubric files, not its top-level prompt. Lemo-Opuscar's DIRECTOR.md self-check guide and per-style demo films are what make results repeatable, and that pattern transfers to any domain where you want an agent to hit a consistent bar.
Related reading
- What are agent skills? The complete guide
- SWE-Game: can coding agents build the games we want?
- Anatomy: a Claude Code skill for isometric interactive figures
- Show Me: a skill for visual output in coding agents
- Chisle: a Claude Code skill for token savings
- Claude Code commands: complete reference
- Official: lemomo-ai/lemo-opuscar on GitHub
Details reflect the repository README as of October 12, 2026; star counts, styles and requirements may change.