English | 中文文档
video-agent-skills
video-agent-skills is a multi-agent team of Claude Code skills that turns a video topic into a DaVinci/CapCut-ready timeline — from research to publish.
Why
Creating a structured knowledge video from scratch requires juggling research, scriptwriting, storyboarding, voice-over, image generation, timeline assembly, and platform metadata — across many tools with no shared context.
video-agent-skills wires up 11 specialized Claude Code skills into a coordinated pipeline. Each skill owns a single role; the producer orchestrates the whole sequence, with human-review checkpoints so you stay in control at every critical decision.
The pipeline outputs a DaVinci Resolve project or a CapCut/JianYing draft — your choice — ready to render.
Features
Pipeline skills (10 skills — run in sequence)
| Skill | Role | Key output |
|---|---|---|
video-agent-producer |
Director / orchestrator. Accepts a topic, breaks it into sub-tasks, dispatches all other agents in order, manages 4 human checkpoints and checkpoint recovery. | project.json + final deliverable package |
video-agent-operator |
Analytics strategist. Reads Douyin account CSV/Excel exports (up to 3 accounts: growth, AI, vlog), identifies high-performing content patterns, and suggests topics. | Analytics report + topic recommendations |
video-agent-researcher |
Research lead. Two modes: search mode (web articles + YouTube transcripts → structured outline) and organize mode (user-supplied material → narrative-restructured outline). | outline.md + materials/sources.json |
video-agent-writer |
Screenwriter. Turns the outline into a word-for-word narration script. Supports 3 opening patterns (counter-intuitive / story / question), 280 chars/min pacing, and citation tagging for downstream agents. | script.md |
video-agent-storyboarder |
Storyboard designer. Calls generate_storyboard.py which sends the script to Gemini Flash (gemini-2.5-flash) and returns a shot-by-shot storyboard with visual descriptions, asset type, and mood per shot. |
storyboard.json + storyboard.md |
video-agent-voice |
Voice producer. Converts the script to audio and subtitles. Supports MiniMax Speech-02 (Chinese, recommended), Edge-TTS (free draft), ElevenLabs (English), and manual recording mode. Reads storyboard mood annotations for emotion control. | audio/voiceover.mp3 + audio/subtitles.srt |
video-agent-visual |
Art director. Reads storyboard.json and calls generate_images.py to batch-generate images via Gemini Image API (Nano Banana). Skips shots flagged as post-production. Supports style presets (default / tech / knowledge). |
visuals/*.png + visual-timeline.json |
video-agent-editor |
DaVinci editor. Imports all assets into DaVinci Resolve Studio via the Python API. Builds a multi-track timeline (V1 visuals, V2 text effects, V3 data effects, A1 voice-over, A2 BGM placeholder, subtitle track). Generates placeholder clips for post-production shots. | DaVinci Resolve project + editor-report.md |
video-agent-jianying-editor |
CapCut/JianYing editor. Shares the exact same inputs as the DaVinci editor but outputs a JianYing draft folder via VectCutAPI. The producer dispatches one or both editors via the output_target parameter. |
jianying-draft/ + jianying-editor-report.md |
video-agent-publisher |
Publishing coordinator. Generates platform-specific metadata (title candidates, description, hashtags, BGM suggestions, citation list) for Douyin, Bilibili, and YouTube. | publish/metadata.json + publish/sources.md |
Independent skill (1 skill — standalone branch)
| Skill | Role | Key output |
|---|---|---|
live-sharing-writer |
Live / presentation scriptwriter. For livestreams, talks, or recorded speeches — not videos. Works by interviewing the presenter (no external research), extracts personal stories and viewpoints, and writes a spoken-word script with interaction cues and pacing markers. | script.md (conversational, speaker-ready) |
Why is
live-sharing-writerseparate? The video pipeline (researcher → writer → storyboarder → …) assumes external research and visual production. Live sharing is presenter-first: all material comes from the speaker, no storyboard or image generation is needed. The two branches share a script format but nothing else.
Install
These are bare Claude Code skills — plain directories, each containing a SKILL.md that Claude Code reads as agent instructions.
Option A — Copy manually
Copy each skill directory into your Claude Code skills folder (the path depends on how you configured Claude Code; typically ~/.claude/skills/ or a project-level .claude/skills/):
cp -r /path/to/video-agent-skills/video-agent-producer ~/.claude/skills/
cp -r /path/to/video-agent-skills/video-agent-researcher ~/.claude/skills/
# … repeat for each skill you need
Option B — Clone the repo next to your project and add to your CLAUDE.md
git clone https://github.com/chenyuxiaojin/video-agent-skills.git
Then in your project's CLAUDE.md (or ~/.claude/CLAUDE.md) reference the cloned path so Claude Code can find the skills.
No marketplace.json or plugin registry — this repo ships raw skill directories intentionally. Install only what you need.
Usage
Start the full pipeline by telling Claude Code:
开始制作视频 — 主题:[your topic]
Or trigger individual skills directly:
# Topic research only
搜索关于 [主题] 的素材
# Analytics + topic suggestion
分析数据,下期做什么选题
# Live sharing script (independent)
帮我准备一个直播分享稿
The producer skill (video-agent-producer) orchestrates all pipeline steps and pauses at 4 checkpoints for your review before proceeding.
Compared to alternatives
| video-agent-skills | xiaochen-skills (cyxj-ai-weekly-news) | faceless-video / auto-shorts generators | Anthropic official agent-skills examples | |
|---|---|---|---|---|
| Primary use | Long-form knowledge video (6–15 min), full pipeline from research to timeline | AI weekly news digest → Douyin short clip (topic discovery + script only, no full pipeline) | Fully automated faceless short videos, minimal human control | Reference implementations for agent patterns, not production-ready |
| Human checkpoints | 4 explicit review gates | None (automated) | None | N/A |
| Editor target | DaVinci Resolve Studio or CapCut/JianYing (your choice) | Not included | Platform-specific auto-upload | Not included |
| AI dependencies | Gemini Flash (storyboard) + Gemini Image (visuals) + MiniMax/Edge-TTS/ElevenLabs (voice) | Grok search + Claude | Varies by tool | Varies |
| Overlap note | — | cyxj-ai-weekly-news covers AI topic discovery for short clips; video-agent-skills handles the full production chain for longer content |
Different audience and length target | — |
FAQ
How many skills are there? 11 total: 10 pipeline skills (producer → operator → researcher → writer → storyboarder → voice → visual → editor → jianying-editor → publisher) plus 1 independent skill (live-sharing-writer).
Do I have to use DaVinci Resolve?
No. You can use CapCut/JianYing instead — just tell the producer output_target: "jianying". You can also generate both (output_target: "both"). The DaVinci path requires DaVinci Resolve Studio to be running; the JianYing path requires VectCutAPI running locally.
How do I install these skills? Copy the skill directories you need into your Claude Code skills folder (see Install above). There is no package manager step.
What APIs do I need?
| API | Used by | Required? |
|---|---|---|
Gemini Flash (gemini-2.5-flash) |
storyboarder (via generate_storyboard.py) |
Yes for storyboard generation |
| Gemini Image API | visual (via generate_images.py) |
Yes for AI image generation |
| MiniMax Speech-02 | voice (Chinese TTS, recommended) | No — Edge-TTS is free alternative |
| ElevenLabs | voice (English TTS) | No — only for English projects |
| Edge-TTS | voice (free draft) | No — bundled, no key needed |
| VectCutAPI (local) | jianying-editor | Only if using CapCut/JianYing output |
| DaVinci Resolve Studio Python API | editor | Only if using DaVinci output |
| Web search / YouTube transcripts | researcher | Requires Claude Code web tools |
Can I run a single skill without the full pipeline?
Yes. Every skill can be triggered independently with a natural-language phrase (see each skill's trigger conditions in its SKILL.md). The producer is optional — it just automates the hand-offs.
License
MIT © 2026 chenyuxiaojin
No comments yet
Be the first to share your take.