cinecode: Claude Code makes short films with no video AI, just code

What it's about

In early November 2026, Turkish developer Sarp Yildiz (@sarpstar on X) released a tool called cinecode. It's a skill for Claude Code, Anthropic's tool that lets Claude work directly on your machine: creating files, running programs and checking the results. A skill is a bundle of instructions and tools that teaches the AI how to handle a specific task. Think of it as a recipe book plus a set of specialist utensils that you hand to a good chef. cinecode also works in the AI code editor Cursor.

Yildiz says you can use it to make short films, YouTube videos and Instagram Reels without a single generative video model. The idea behind it is simple. The AI doesn't paint the video itself. It writes a program that draws it. The script, scenes, sound and edit all end up as code, and the output is a finished MP4 file. cinecode is released under the MIT license, so anyone can use it for free, modify it and use it commercially. The code is on GitHub, and the docs are available in English and Japanese, among other languages. This article draws on a detailed analysis on the blogging platform note.com, which checks the developer's claims against the official documentation.

How the tech works

The process has three steps:

1. Claude writes a program that draws the video frame by frame.
2. The program runs in a headless browser, a copy of Google Chrome working invisibly in the background. There it plays back and gets captured one frame at a time.
3. The standard tool ffmpeg combines those frames with the audio, which is also generated by code, into an MP4 file.

According to the official description, every frame of the finished video is drawn from code in the browser. For 3D scenes, cinecode uses three.js, a popular library for 3D graphics on the web.

Here's the key difference from video AIs like the well-known text-to-video models. Give those models the same prompt and you'll often get a different result every time, so you keep rerolling until something works. Chinese-language posts about cinecode compared this to a gacha game, the loot-box mechanic from mobile games where you never know what you'll get. cinecode instead describes a video like a mathematical formula: put in a point in time and you get exactly the frame for that moment. The same formula produces exactly the same video on every run. The technical term is deterministic.

In practice, that has clear advantages:

• If you want "just the text at second 3 a bit bigger," you change exactly that spot in the code.
• Everything else stays the same, so a fix is less likely to break something elsewhere.
• You can keep old versions of the code and restore them at any time.

The author of the analysis puts it this way: a video AI is like a portrait redrawn by hand every time. cinecode is more like a model kit built from a blueprint. Change one line in the plan and only that one part gets rebuilt.

Using it: answer questions instead of writing code

Yildiz says using it feels "like answering fun multiple-choice questions." According to the manual, there are five stages:

1. Pitch your idea: One sentence is enough. If you already have a detailed concept, you go straight to production.
2. Answer follow-up questions: These cover how to interpret the topic, visual style (3D, hand-drawn, paper look and more), length and aspect ratio, mood, and sound and music. For each question, the AI marks its recommendation with a star (★).
3. Approve the plan: You get a timeline of what happens at each second, plus "director's notes" on camera work and color grading. You can edit both line by line. The AI then shows you two scenes as stills so you can check the mood before the full video gets made.
4. Review the prototype: First comes a low-resolution version that the AI checks for quality automatically. On top of that, a second AI that knows nothing about how the piece was made watches it like a first-time viewer and gives its verdict. Fixes are applied only where you ask for them.
5. Get the final version: A full-quality export with balanced audio levels.

There are two ways to work. Quick Mode puts everything into a single file and is meant for videos under about three minutes, like social media clips or ads. It has roughly three approval steps. In Film Mode you sign off on each stage separately: outline, script, sample frame, rough cut (a simple version for checking motion and timing) and final production. It's designed for researched documentaries and pieces over ten minutes. You can start in Quick Mode and switch to Film Mode later.

What you can make with it

The manual lists these use cases:

• Short films: cinecode randomly pairs ideas with real things like museum pieces, animals or space photos to find unusual angles. It checks your earlier work so nothing gets repeated.
• YouTube videos over ten minutes: Documentaries and explainers structured in acts or chapters. Claims about real people and events come with sources. Photos, maps and historical footage are pulled from public archives.
• Ads and product videos: Your brand's colors, fonts, logo and closing call to action get applied in three formats at once: vertical (9:16), square (1:1) and landscape (16:9).
• Vertical explainers: 30 to 60 seconds long, with a question in the very first second.
• Music videos: cinecode measures a song's rhythm and times scene changes to bars and beats.
• Videos based on a reference: Show cinecode a video you like and it measures only the format: shot length, pacing and use of color. The visuals and sound are then created from scratch. No original footage or music is reused.

Formats that work well can be saved as templates and reused for whole series.

Is it really "just a Claude subscription, no extra costs"?

Yildiz's boldest claim is that a Claude subscription is all you need and you won't pay for anything else. According to the analysis, that's broadly true, with some conditions:

• Only in Claude Code or Cursor. The manual explicitly says cinecode doesn't work in the regular chat on claude.ai, because it can't run tools on your own machine there. So you need a plan that includes Claude Code.
• You need your own computer. Requirements are Node.js 22 or later, Python 3.9 or later, ffmpeg, and Google Chrome or a similar browser. The AI will install anything missing after asking you. Rendering needs a graphics card the browser can use. Current Macs or Windows PCs with working graphics drivers are fine.
• Narration goes through other services. You don't necessarily have to pay for it. Options include your own Google Gemini key (with a free tier), the voice service ElevenLabs, or a free Turkish text-to-speech engine. The suggested voices are mostly Turkish and English.
• Usage is heavy. According to the manual, a 30- to 60-second video takes the AI around 100 to 250 steps. Claude plans have usage limits, so longer projects hit them sooner. cinecode tries to compensate, for example by suggesting a fresh chat when the current one gets too long.
• "No video models at all" comes with caveats. The longer and more fact-heavy a piece is, the more photos and videos from public archives go into it, from sources like Wikimedia Commons, the Internet Archive, NASA or museums. If you connect your own image or video AI, it can serve as a last resort. According to the docs, though, the AI won't suggest this on its own.

Copyright built in from the start

For companies, rights are often the biggest hurdle. cinecode bakes them in. It only uses fonts licensed for commercial use without attribution. External images and sounds come only from the public domain, meaning no rights apply, or under CC0, where the creator has waived their rights. The license is recorded at import. Every video comes with a credits file you can paste straight into the video description. It's also meant to help if YouTube asks about music rights. Copyrighted works collected as references are used for orientation only and never end up in the video. But you're responsible for any logos, footage or music you bring yourself. The software is provided "as is" and is explicitly not legal advice.

Background: video from code is catching on

cinecode is part of a trend that picked up speed in 2026. In January, Remotion, a tool for making videos programmatically, released an official skill for Claude Code. According to reports, its demo video got more than six million views and the skill was installed over 25,000 times in its first week. After Claude Opus 5.5 launched on September 22, X filled up with videos supposedly made "with just prompts." Usually they were powered by programs drawing frame by frame, plus a lot of work. One creator said a short watercolor-style video took about seven hours and 163 prompts.

cinecode itself spread fast. On November 7, the project had 518 stars and 87 forks on GitHub, meaning copies that other people develop further. There's already a port for OpenAI's coding AI Codex and a volunteer-built desktop app called "Cinecode Studio." The video service animai has added a template called "Cinecode Film" to its lineup and charges 180 credits per minute for it. Within days, a free method became another company's paid feature, which the MIT license explicitly allows.

Our take

cinecode isn't a new AI model. It's a clever workflow that plays to what today's language models do well: writing code. If you need explainers, product clips in multiple formats or recurring series, it gives you something video AIs have barely offered so far: consistent results, targeted fixes and clean documentation of rights. The author of the analysis sees this as a shift in video production. You pay less for manual work like shooting, drawing and editing, and more for judgment: deciding what to pick and what to fix. And because open-source methods get copied quickly, the real edge lies in accumulated templates, brand assets and client relationships.

There are limits, though. Real actors, facial expressions and photorealistic live-action scenes are hard to pull off with code. That's where cameras and video AI keep the upper hand. Visual quality depends heavily on what browser graphics and three.js can do. There are no independent tests yet of quality or actual subscription usage. The figures come from the project documentation and from the developer himself. And "no extra costs" assumes a paid Claude subscription, which can hit its limits quickly at 100 to 250 steps per short video.