skill

vision-art-creator

AI film-production skill pack for Claude Code and OpenAI Codex: 11 creator-* skills from screenplay to final cut, with a recorded production graph. Docs in 12 languages.

ilkaydemiralay/vision_art_creatorcommunitylistedunreviewednot running here yet

Use skill

Set up

Skills are installed by placing their folder where your agent loads skills.

Install instructions come from the README

Skills do not use an MCP config. Follow the steps in the README below; one-click install arrives with accounts.

README

synced Oct 14, 2026

vision_art_creator

English · 中文 · Español · हिन्दी · العربية · Português (BR) · Türkçe · Français · Deutsch · Русский · 日本語 · 한국어

An AI film‑production skill pack for Claude Code and OpenAI Codex.

vision_art_creator bundles 11 creator-* skills that cover every department of a film production — from the screenplay all the way to the final cut — in a single repository. Install it on any machine with git clone + ./install.sh.

The skills are designed to reference one another (creator-pipeline-supervisor orchestrates the rest). Installing them all together is recommended.

Demo: Before She Leaves

A 26-second scene produced with this pack and Higgsfield (Nano Banana Pro references, Seedance 2.0 clips), plus a 30-second making-of that shows the recorded graph: owners, hashes and approvals. Two agents, Claude Code and OpenAI Codex, worked from the same skill files. All footage is AI-generated.

Film (26 s)

https://github.com/user-attachments/assets/1af79d83-494b-403b-96df-cfdf44dabbf1

Making-of (30 s)

https://github.com/user-attachments/assets/5db19cde-1c08-42f1-a48a-cd3a4a0a28d8

Full-quality files: release v1.2.0

Installation

git clone https://github.com/ilkaydemiralay/vision_art_creator.git ~/projects/vision_art_creator
cd ~/projects/vision_art_creator
./install.sh

Using OpenAI Codex? Install into its skills directory instead, then restart Codex:

./install.sh --codex   # ~/.agents/skills/

install.sh creates a symlink for each skill under ~/.claude/skills/<skill-name> that points back into this repo. The benefit: to update, a simple git pull is enough — no reinstall required.

Options

./install.sh --target /path/to/skills   # install into a different skills dir
./install.sh --force                    # overwrite existing names
./uninstall.sh                          # remove the symlinks

uninstall.sh only removes symlinks that point into this repo — it leaves foreign links and real directories untouched (unless --force).

Verify

After installing, restart Claude Code and type:

/creator-pipeline-supervisor

Check that all 11 creator-* skills appear in the skill list.

What's in the pack

Skill	Summary
`creator-pipeline-supervisor`	Orchestrates the whole production, sequences the departments, enforces continuity, runs QC, and produces the delivery‑readiness report.
`creator-director`	Translates the screenplay into a unified directorial vision: scene direction, performance, blocking, tonal control.
`creator-screenwriter`	Writing and revising screenplays, treatments, loglines, scene outlines, and dialogue.
`creator-character-designer`	Designs the character as an integrated whole: psychology, biography, visual identity, costume, props, FACS‑coded expressions.
`creator-production-designer`	Builds the world of the film: locations, sets, props, period atmosphere, color/material language, continuity anchors.
`creator-cinematographer`	Designs the visual language: light, camera, lens, framing, color, atmosphere, movement.
`creator-storyboard-artist`	Visualizes scenes panel by panel: shot scales, angles, blocking, composition, AI prompts.
`creator-shot-list-designer`	Turns scenes and storyboards into a technical shot list, broken into AI‑producible chunks.
`creator-sound-music-designer`	The sonic world of the film: atmosphere, foley, SFX, score, leitmotifs, scene‑by‑scene music plan, AI audio prompts.
`creator-prompt-engineer`	Converts every department's output into consistent prompts for GPT Image 2.0, Nano Banana, Sora, Veo, Runway, Kling, Higgsfield, and more.
`creator-final-cut-editor`	Assembles AI‑generated shots/audio/music/graphics into a finished film: rough/fine/final cut, AI‑error triage, delivery formats.

The full definition of each skill lives in its own SKILL.md file.

How it works

The pack runs on filesystem‑based shared state. All skills read from and write to a common project/ tree (bible/, screenplay/, characters/, storyboards/, prompts/, cuts/, qc/, …). creator-pipeline-supervisor maintains the canonical files (the project & continuity "bibles") and audits every department's output against them.

The canonical pipeline:

0. project bible & vision
1. creator-screenwriter        → screenplay
2. creator-director            → vision, direction sheets, arcs
3-4-5. creator-character-designer + creator-production-designer
        + creator-cinematographer        (run in parallel)
6. creator-storyboard-artist   → panels with prompts
7. creator-shot-list-designer  → shot list + edit plan
8. creator-prompt-engineer     → tool-fit image + video prompts
   → [AI material generation — operator]
9. creator-sound-music-designer → sound + score plan
10. creator-final-cut-editor   → rough → fine → final cut → delivery

Throughout: creator-pipeline-supervisor enforces continuity, runs QC,
manages revision loops, and holds the bibles.

The sequence is canonical but not rigid: director feedback can re‑trigger the screenwriter, and character/production/cinematography typically run in parallel once the directorial vision is set.

Recorded graph mode (optional, experimental)

Since v1.1.0 the pack includes an optional, machine-checkable workflow for a single scene. creator-pipeline-supervisor can run pre-production as a 12-node graph: every artifact is recorded with its SHA-256 hash, every approval is bound to the exact inputs it reviewed, and a revision re-runs only the nodes it affects.

Workflow: workflows/single-scene.v1.json. Contract: skills/creator-pipeline-supervisor/references/graph-workflow.md.
JSON Schemas in schemas/, a read-only validator in scripts/validate_graph.py, tests in tests/.
A worked 15-second, three-shot text pilot in examples/single-scene/: v00 stops at a design conflict, v01 resolves it, v02 changes one prop colour.

Normal skill use needs no Python. To run the validator (Python 3.10+):

python3 -m venv .venv
.venv/bin/pip install -r requirements-graph.txt
.venv/bin/python -m unittest discover -s tests
.venv/bin/python scripts/validate_graph.py validate --manifest path/to/manifest.json

Limits: this is a text-only pilot written by a single author. Parallel department runs, media generation, editing and cost were not tested. Approver identity and timestamps are declared by the operator, not signed. Design and results notes are in docs/ (Turkish).

Updating

cd ~/projects/vision_art_creator
git pull

Because the skills are symlinked, no extra step is needed.

See CHANGELOG.md for what changed in each release. v1.1.0: creator-cinematographer and creator-storyboard-artist now keep prompt drafts in their own folders; only creator-prompt-engineer writes final prompts under project/prompts/.

Development

Edit a skill in the repo (skills/creator-*/SKILL.md).
Test the change in Claude Code — since it's a symlink, it takes effect immediately.
Commit + push.

To add a new creator skill:

mkdir -p skills/creator-new-skill
# write SKILL.md and README.md
./install.sh   # create the symlink for the new skill

Translations

This README and each skill's README.md are available in 12 languages (see the language selector at the top). The SKILL.md instruction files are kept in English on purpose — Claude responds in the user's language at runtime, and a single canonical instruction set avoids duplicate skill names.