Create Storyboard Skill

Create continuity-first storyboard production packages for Image 2 and SceneDance/Seedance video generation.

This skill turns a script, ad idea, short-drama concept, product video brief, or scene outline into a director-grade production package: bibles, shot cards, clip plans, handoff matrices, edit boundary plans, Image 2 prompts, SceneDance prompts, CapCut/Jianying edit notes, and image manifests.

What It Produces

• Script analysis and project brief.
• Character, scene, prop/product, style, and continuity bibles.
• Shot cards with one SH### per default CLIP###.
• SceneDance reference input matrix.
• Handoff design and edit boundary matrices.
• Chinese and English Image 2 prompts.
• SceneDance shot prompts and usage list.
• Post-edit, Jianying/CapCut, risk, fallback, and image manifest files.

Reproducibility Profile

The scaffold script is deterministic and does not require an AI model. Production-quality storyboard planning does require a capable agent runtime because the hard work is continuity design, shot handoff planning, visual reference discipline, prompt writing, and QA.

Recommended agent runtime:

• Codex-style agent mode with local file read/write and command execution.
• Multimodal model when images, character sheets, keyframes, or storyboard boards are part of the task.
• Frontier reasoning model, such as GPT-5.5 or an equivalent model, for long-form drama, complex continuity, or high-value production work.
• Reasoning effort: high for normal production work; xhigh when available for multi-scene continuity, dense handoff design, or bilingual prompt packages.
• Long enough context to inspect script briefs, bibles, shot cards, reference matrices, generated images, and edit notes together.

Known-good author setup: macOS, Codex-style local agent, GPT-5.5-class reasoning, and xhigh reasoning for complex packages. Smaller models can still use the scaffold, but may need more human correction for continuity, handoff logic, and prompt consistency.

Quick Start

Create a package skeleton from the repository root:

python3 skills/create-storyboard/scripts/create_storyboard_package.py demo-product-ad \
  --root outputs \
  --title "Demo Product Ad" \
  --duration "30s" \
  --aspect "16:9" \
  --video-type "product ad"

The command writes a complete template package under:

outputs/storyboard_projects/demo-product-ad/

Platform Notes

The repository is authored and validated primarily on macOS. The scaffold script uses only the Python standard library and is cross-platform.

• macOS/Linux/WSL2: use the commands as written.
• Windows PowerShell: use py instead of python3, and use backticks or single-line commands instead of POSIX backslash continuations.
• Python 3.10+ is recommended.
• If the workflow also includes image generation, video generation, or local media tooling, validate those tools separately on the target OS.

Skill

The reusable agent skill lives at:

skills/create-storyboard/SKILL.md

For Codex-style local installation, copy skills/create-storyboard/ into your local skills directory and restart the agent.

The installed skill folder is self-contained for a smoke test:

cd ~/.codex/skills/create-storyboard
python3 scripts/create_storyboard_package.py demo-storyboard \
  --root outputs \
  --title "Demo Storyboard" \
  --duration "30s" \
  --aspect "16:9"

Read skills/create-storyboard/references/runtime-and-platform.md for agent runtime and OS notes.

Design Principle

The important boundary is:

script or concept -> continuity bibles -> shot cards -> handoff/edit matrices -> prompts -> generated images/video -> edit plan

Storyboard boards are review and reference artifacts. Clean keyframes are the preferred primary inputs for SceneDance/Seedance. Final production boards should be assembled deterministically from clean panels/keyframes and project metadata, not accepted as raw AI-generated board layouts.

Status

This repository is an open-source packaging pass over a working local Codex skill. The deterministic scaffold path is included; image and video generation remain dependent on the user's chosen tools and model/runtime access.