cinematic-skill draws documentary films in code
An MIT-licensed agent skill turns a one-line request into a narrated, code-rendered documentary with timed captions and sound.
A film pipeline shipped as an agent skill
On 2026-11-09 a GitHub account named Moh4696 published cinematic-skill, an agent skill for narrated cinematic films built entirely from code. Five days later, on 2026-11-14, it had 312 stars and 21 forks. The repository is Python, MIT licensed with exploraX named as copyright holder, and tagged agent-skills, claude-code, codex, motion-graphics and video.
Its own example: ask your coding agent for "a 90 second documentary about the Wright brothers" and it writes the script, records a voice, finds public-domain photos with their licences, and builds one continuous camera move over a single designed world. Stamps land, highlighters swipe, counters roll up on the exact word the narrator says, and the film finishes with sound effects, an optional score and burned-in captions.
Every frame is drawn, none is generated
The cinematic-skill repository states plainly that nothing is generated video. Skia draws each frame from Python, so a render comes out identical every time, edits are cheap, and a new voice re-times the whole film by itself. The bundled worked example, a 90 second documentary called "Twelve Seconds", is built from seven public-domain photographs.
That is a real trade-off against prompt-to-video tools: you get a scene file rather than a clip, which is repeatable and editable, but the look is limited to the elements the skill provides.
How a film gets built
The repo splits production into six stages.
- Story. The agent researches, fact-checks and writes one idea per line, each with a camera note. A beat table with a chronology check comes to you for approval.
- Voice. A free scratch voice (macOS
sayor espeak-ng) times the film first. The final voice can be any TTS or a person, split into timed lines, with suspect cuts flagged. - Images. Wikimedia Commons search and download with every licence recorded; non-free files are refused. Your own photos and AI stills are allowed, and a credits file is written.
- Scene. A short Python file places items against the voice timing, using calls such as
doc.W(line, "word"), and keys the camera. - Review. The agent renders stills and contact strips of each move and checks them before showing you a draft.
- Finish. Sound effects are synthesised, mixed with the voice and an optional score, ducked and loudness-normalised, then rendered in parallel with four caption styles and an SRT.
Install and the first render
Install is npx skills add Moh4696/cinematic-skill on Node 22.20 or newer, which the repo says covers Claude Code, Codex, Cursor, Gemini CLI, OpenCode and others; start a new agent session afterwards. Claude Code and Codex also have plugin commands, Gemini CLI an extension install, and a shell installer clones to ~/.cinematic-skill for machines without Node. The scripts need pip install -r requirements.txt (skia-python, numpy, pillow, scipy optional) and ffmpeg on PATH; check_setup.py reports what is missing. On a 16 GB laptop, the repo says the 92 second example renders at 1080p in about six minutes.
Looks, parts and open questions
Four presets are documented: documentary with a cork evidence wall and black and white archive, product launch in studio charcoal, science and engineering on blueprint paper, and a night look for tech and space. Placeable items include photos, torn clippings, typeset documents, cards, distressed type, rolling counters, stamps, pen circles, highlighter swipes, red string and maps with filled regions, routes and markers.
What the source leaves open: how long a full production takes end to end, whether non-English narration is supported, what happens with music licensing, and whether the six minute render holds at higher resolutions. Install is not zero-setup, since Node or a shell installer, Python packages and ffmpeg are all required.
What this suggests for AI-assisted builds
An agent skill of this shape, a bundle of scripts, rules and approval checkpoints rather than one long prompt, is one way to make agent output repeatable. Where a visual has to follow a copy change, re-rendering code beats re-rolling a video model. If you are building an app around generated content, the same discipline helps: describe the outcome precisely and let the tool work from a fixed set of parts, which is what writing effective prompts in Draftship covers for app builds.
Primary sources
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.