konlu/story-studio

Story Studio is a skill package for AI agents such as Claude Code, Codex, or Cursor. It carries a short-video production pipeline from topic selection and scriptwriting through voiceover, storyboarding, illustration, and finally an editable JianYing Pro draft, aimed at history and real-person narrative channels.

What it does

Rather than generating a finished video file, the package routes an agent through four modules: scriptwriting with two built-in writing styles, storyboard and image generation with three visual styles, TTS and subtitle alignment, and export to a JianYing draft. The writing module bundles two distilled styles: historical narrative, drawn from 57 transcripts, and character story, drawn from 300+ transcripts. Both include hook templates, topic strategies, and fact-checking rules README. Outputs stay in a user workspace project folder as script, narration audio, SRT, storyboard Markdown/JSON, images, covers, and the draft project.

How the pipeline works

The distinctive mechanism is audio-first storyboarding: instead of splitting the script evenly, the agent aligns shot cuts to real speech timestamps, so the first sentence has a matching frame at frame zero README. Voiceover has two paths. Path A synthesizes narration through Doubao speech (V3 bidirectional streaming WebSocket), requiring an API Key, speaker ID, and resource_id from the Volcano Engine console; the key is stored via environment variable or OS-level encrypted credential storage, not in the project README. Path B accepts a user-supplied MP3. Either way, local Coli (based on sherpa-onnx) produces the timestamped SRT and cross-checks names and proper nouns against the final script.

Export uses pyJianYingDraft to write the JianYing Pro draft, copying media into the draft directory and adding slow push/pull keyframes plus punch-zoom cuts at transitions. Optional local FFmpeg rendering can output final.mp4 without opening JianYing. Covers are 3:4 and 4:3 by default; subtitles are off unless requested.

Flexible scope

Three switches—writing, illustration, and draft—let users run only part of the workflow: just a script and hooks, just storyboard prompts for external image tools, or just arranging existing audio and images into a draft README. The installer provides three routes: ask the agent to clone and install, run a one-line git clone into the agent's skills directory, or run install.bat/install.sh from a downloaded zip.

Requirements and limits

Multimedia dependencies are initialized on demand by the bootstrap scripts into a user data directory: Python 3.11 via uv, Node.js 20+ LTS with Coli, FFmpeg/ffprobe, and pyJianYingDraft; JianYing Pro itself is required for draft work and can be installed via WinGet on Windows README. Writing-only use avoids those components. The skill is free under MIT, but Doubao TTS and any external image-generation service may charge their own fees; the README says the agent asks before first synthesis or batch image generation. The bundled corpus is for learning narrative structure and is not a verified historical source, and the reference images are not historical photographs. Platform support statements center on Windows, macOS, and Linux, with JianYing draft detection and installation specifically noted for Windows.