{"id":"77404807-33d4-4d78-b619-d06a0336c84e","slug":"video-agent-skills","kind":"skill","name":"video-agent-skills","summary":"Multi-agent Claude Code skills that turn a topic into a DaVinci/CapCut-ready video timeline — research to publish.","owner":"chenyuxiaojin","repoUrl":"https://github.com/chenyuxiaojin/video-agent-skills","license":"MIT","licenseTier":"permissive","trust":"community","stars":11,"match":null,"fewStars":false,"format":"skill","createdAt":"2026-02-07T13:54:45.000Z","addedAt":"2026-10-14T06:10:25.092Z","logoUrl":"https://avatars.githubusercontent.com/u/52022516?u=09cfa5476915f6c8b66e8cefca77fcb494f34dc2&v=4","linkOnly":false,"collection":null,"skillPath":null,"categories":["design-media","developer-tools"],"tags":["ai-video","capcut","claude-code","content-automation","davinci-resolve","multi-agent","storyboard","text-to-speech"],"runtime":"binary","local":true,"cloudCapable":false,"cloudCapability":"unknown","cloudReadyOnPlatform":false,"needsApiKey":false,"reviewed":false,"syncedAt":"2026-10-17T09:31:40.836Z","sample":false,"url":"https://conduits.dev/tools/video-agent-skills","language":"Python","transports":["stdio"],"install":{},"requirements":{"env":[],"upstreamMayCharge":false},"tools":null,"readme":"**English** | [中文文档](README.zh-CN.md)\n\n# video-agent-skills\n\n> 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.\n\n[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)\n[![Skills: 11](https://img.shields.io/badge/Skills-11-blue.svg)](.)\n[![Platform: Claude Code](https://img.shields.io/badge/Platform-Claude%20Code-blueviolet.svg)](https://claude.ai/code)\n\n## Why\n\nCreating 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.\n\nvideo-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.\n\nThe pipeline outputs a DaVinci Resolve project or a CapCut/JianYing draft — your choice — ready to render.\n\n## Features\n\n### Pipeline skills (10 skills — run in sequence)\n\n| Skill | Role | Key output |\n|-------|------|-----------|\n| `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 |\n| `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 |\n| `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` |\n| `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` |\n| `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` |\n| `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` |\n| `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` |\n| `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` |\n| `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` |\n| `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` |\n\n### Independent skill (1 skill — standalone branch)\n\n| Skill | Role | Key output |\n|-------|------|-----------|\n| `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) |\n\n> **Why is `live-sharing-writer` separate?**\n> 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.\n\n## Install\n\nThese are bare Claude Code skills — plain directories, each containing a `SKILL.md` that Claude Code reads as agent instructions.\n\n**Option A — Copy manually**\n\nCopy 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/`):\n\n```bash\ncp -r /path/to/video-agent-skills/video-agent-producer ~/.claude/skills/\ncp -r /path/to/video-agent-skills/video-agent-researcher ~/.claude/skills/\n# … repeat for each skill you need\n```\n\n**Option B — Clone the repo next to your project and add to your CLAUDE.md**\n\n```bash\ngit clone https://github.com/chenyuxiaojin/video-agent-skills.git\n```\n\nThen in your project's `CLAUDE.md` (or `~/.claude/CLAUDE.md`) reference the cloned path so Claude Code can find the skills.\n\n**No marketplace.json or plugin registry** — this repo ships raw skill directories intentionally. Install only what you need.\n\n## Usage\n\nStart the full pipeline by telling Claude Code:\n\n```\n开始制作视频 — 主题：[your topic]\n```\n\nOr trigger individual skills directly:\n\n```\n# Topic research only\n搜索关于 [主题] 的素材\n\n# Analytics + topic suggestion\n分析数据，下期做什么选题\n\n# Live sharing script (independent)\n帮我准备一个直播分享稿\n```\n\nThe producer skill (`video-agent-producer`) orchestrates all pipeline steps and pauses at 4 checkpoints for your review before proceeding.\n\n## Compared to alternatives\n\n| | video-agent-skills | [xiaochen-skills](https://github.com/chenyuxiaojin/xiaochen-skills) (cyxj-ai-weekly-news) | faceless-video / auto-shorts generators | Anthropic official agent-skills examples |\n|---|---|---|---|---|\n| 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 |\n| Human checkpoints | 4 explicit review gates | None (automated) | None | N/A |\n| Editor target | DaVinci Resolve Studio **or** CapCut/JianYing (your choice) | Not included | Platform-specific auto-upload | Not included |\n| AI dependencies | Gemini Flash (storyboard) + Gemini Image (visuals) + MiniMax/Edge-TTS/ElevenLabs (voice) | Grok search + Claude | Varies by tool | Varies |\n| 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 | — |\n\n## FAQ\n\n**How many skills are there?**\n11 total: 10 pipeline skills (producer → operator → researcher → writer → storyboarder → voice → visual → editor → jianying-editor → publisher) plus 1 independent skill (live-sharing-writer).\n\n**Do I have to use DaVinci Resolve?**\nNo. 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.\n\n**How do I install these skills?**\nCopy the skill directories you need into your Claude Code skills folder (see [Install](#install) above). There is no package manager step.\n\n**What APIs do I need?**\n| API | Used by | Required? |\n|-----|---------|-----------|\n| Gemini Flash (`gemini-2.5-flash`) | storyboarder (via `generate_storyboard.py`) | Yes for storyboard generation |\n| Gemini Image API | visual (via `generate_images.py`) | Yes for AI image generation |\n| MiniMax Speech-02 | voice (Chinese TTS, recommended) | No — Edge-TTS is free alternative |\n| ElevenLabs | voice (English TTS) | No — only for English projects |\n| Edge-TTS | voice (free draft) | No — bundled, no key needed |\n| VectCutAPI (local) | jianying-editor | Only if using CapCut/JianYing output |\n| DaVinci Resolve Studio Python API | editor | Only if using DaVinci output |\n| Web search / YouTube transcripts | researcher | Requires Claude Code web tools |\n\n**Can I run a single skill without the full pipeline?**\nYes. 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.\n\n## License\n\n[MIT](LICENSE) © 2026 chenyuxiaojin\n"}