Skip to content

About

Film craft and AI video direction skills for Claude Code, Claude Desktop and GitHub Copilot

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

cinewright

A cinema camera on a tripod aimed at a storyboard wall, with a strip of film across the desk

Film craft for AI video. cinewright makes an agent plan a film like a crew and check it like a script supervisor, so separately generated shots keep the same people, wardrobe, props, eyelines, screen direction and light. Each shot is a generator-neutral shot card; a compiler turns a card plus the project's bibles into a given model's prompt.

Status: 0.1.0, the first release. Install routes proved on Windows 11: Claude Code, GitHub Copilot CLI and gh skill.

What is here

Plugin Skills For
cinewright cinewright (director and router), cinewright-shots (shot list and coverage), cinewright-continuity (script supervisor), cinewright-genvideo (prompt compiler for nine models), cinewright-qc (take review with ffmpeg) everyone
cinewright-craft cinewright-script (logline, beats, dialogue), cinewright-design (turnarounds, props, costume, color), cinewright-sets (set plans, wall strings, plates), cinewright-movement (weight, fights, battles, hard subjects), cinewright-camera (lens, light, format), cinewright-history (style cards), cinewright-edit (cutting, pacing), cinewright-finish (grade, crop, VFX, delivery), cinewright-sound (layers, mix, loudness), cinewright-voice (character voices that match shot to shot) deeper craft
cinewright-dev cinewright-curate (add, verify and retire knowledge-base entries; refresh model cards; works in a clone of this repository) maintainers

A worked example lives in examples/three-shot/: a one-line idea taken through brief, script, shot list, bibles, design notes, a continuity diff that catches a planted error, and compiled Veo and MiniMax H3 prompts.

Install

Host How
Claude Code /plugin marketplace add m4bwav/cinewright, then /plugin install cinewright@cinewright (and cinewright-craft@cinewright for the craft skills)
GitHub Copilot CLI copilot plugin marketplace add m4bwav/cinewright, then copilot plugin install cinewright@cinewright
Any agent with the GitHub CLI gh skill install m4bwav/cinewright cinewright-continuity --agent claude-code (plain skill names; --agent picks the host folder)
claude.ai Customize > Skills > Upload skill, one ZIP per skill from the release assets

Install the core plugin first: the craft skills compile through cinewright-genvideo's model cards. Installing copies the skill folders; nothing runs until a skill is used.

What it runs and fetches

Each skill ships scripts/cine.py (Python 3.9+, standard library only). It reads and writes JSON and Markdown in your project folder. It makes no network calls and never renders: writing a prompt never starts a paid render. The qc commands call ffmpeg and ffprobe on clips you already have.

Developing

python -m unittest discover -s tests     # tests
python scripts/cine.py kb lint           # schemas, links, copies, manifests
python scripts/cine.py budget            # size budgets (green, yellow, red)
python scripts/cine.py build             # copy shared/ into skills, rebuild indexes
python scripts/cine.py zip cinewright-continuity   # claude.ai upload ZIP in dist/
python scripts/cine.py kb due                      # entries past their re-check interval (cinewright-curate)
python scripts/cine.py scrub --names FILE          # private details in tracked files, before going public

Edit shared vocabulary, schemas and the library in shared/, never the copies inside skills; build copies them and kb lint fails on drift. The plan is ai-docs/plans/PLAN.md.

Credits

Ideas (not text) from MIT-licensed skill repositories, including DirectorSKILL: response-size ceilings, the repair cost ladder, one owning reference per dimension, continuing from the previous take's observed end state, one change per reroll. Film craft from Arijon, Block, Bordwell and Thompson, Brown, Cousins, Field, Katz, Landis, LoBrutto, Mascelli, McKee, Miller, Murch, Riley, Rowlands, Snyder, Thomas and Johnston, and Weston, cited in each entry.

Banner image generated locally with Z-Image Turbo in ComfyUI, seed 18210612; prompt and workflow in assets/banner-workflow.api.json.

License

MIT, see LICENSE.

About

Film craft and AI video direction skills for Claude Code, Claude Desktop and GitHub Copilot

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages