制作产品视频的技能
Skill to Create Product Videos

原始链接: https://github.com/aaitor/skill-video-studio

skill-video-studio 是一个由 Claude Code skill 编排的本地视频制作工作区。它使用 Playwright CLI 捕获真实的产品交互,使用 HyperFrames 构建品牌化的标题/章节/CTA 场景,并使用 FFmpeg 主控与检查输出。从静默演示开始,然后可选添加旁白、音乐、字幕或主持人画面。仓库包含两个示例项目、可复用的品牌/简介/模板,以及通过 brief.yaml 进行配置,另有用于旁白、头像和音乐的可选环境变量。生产环境需要 Node 22+、FFmpeg/ffprobe、Chromium 和辅助 Python 依赖;完整渲染和 GPU 功能为本地/手动。工作流为:brief → explore → brand → capture → compose → 可选层 → render/QA → thumbnail。私有素材和凭证保持在 Git 之外。文档涵盖入门、使用、配置、故障排除和许可。MIT 适用于原创内容;嵌入字体和外部工具保留各自的条款。

一位开发者开源了一个 Claude Code 技能,用于自动创建带配音和字幕的产品演示视频。他们指出,自己(以及他人)往往更偏向开发而忽视推广。该帖子引发了褒贬不一的反应: - 一些人称赞这个想法,但批评了演示的音频(旁白被截断、音乐过响、混音差)。 - 另一些人认为,像 Opus 这样的现代模型即使不使用技能也能很好地生成视频;一项对比显示,使用技能和不使用技能的结果几乎相同,但技能消耗了更多 token。 - 多位评论者分享了使用 Claude Code、ffmpeg、Playwright 以及 ElevenLabs 等配音工具的类似工作流。 - 有人警告 Playwright 存在 CVE 漏洞。 - 一个反复出现的主题是:分发和营销仍然是真正的挑战,有些人使用 AI 寻找相关论坛发帖,而不是完全自动化内容。 总体而言,社区看到了自动化演示视频的价值,但质疑专用技能相比基础模型能力能增加多少。
相关文章

原文

Video Studio Skill — from product workflow to polished demo

From product workflow to polished demo.

v1.0.1 release · aaitor/skill-video-studio

Quick start · Worked videos · Usage · Contribute

MIT license Offline source and documentation checks

A local video-production workspace driven by the /devrel-video Claude Code skill. Capture real product interactions with Playwright CLI, compose title cards, chapters, and CTAs with HyperFrames, then master and inspect the result with FFmpeg. Start with a silent walkthrough; add narration, music, subtitles, or your own presenter footage when needed.

Install Video Studio Skill and prepare a real walkthrough brief in Claude Code

A real Claude Code session using the installed project skill: clone the studio, prepare an Acme walkthrough brief, and inspect the three-beat plan on a black background. This short demo stops before capture and rendering; it does not claim a video was produced. The recording guide includes the actual generated brief and reproduction command.

For planning: Git and authenticated Claude Code. Clone the full workspace so its native skill and project layout stay together:

git clone https://github.com/aaitor/skill-video-studio.git
cd skill-video-studio
claude

Ask Claude:

/devrel-video Prepare a 25-second silent walkthrough brief for the bundled
Acme app: show the workspace, filter by category, then open a project.
Reuse brands/acme. Show the plan before capture or rendering.

Expected result: a project brief and a concrete beat plan. For a complete video, install the capture/render prerequisites and follow the worked walkthrough. The production stack needs Node.js 22+, FFmpeg/ffprobe, Chromium, and the helper-specific Python dependencies. Optional narration and avatars require your own additional services or GPU setup.

  • Real product capture: browser interaction scripts with explicit beats and stable selectors.
  • Branded composition: title, browser frame, chapters, and CTA in a HyperFrames timeline.
  • Local mastering and QA: master/web copies, ffprobe checks, contact sheet, and thumbnail.
  • Optional audio: six local music styles, narration-driven timing, and voice variants.
  • Optional presentation: subtitles, private bring-your-own presenter footage, and terminal/browser sequences.
  • Source-based workflow: reusable brands, briefs, templates, and two complete worked examples.

Portrait/social output profiles remain on the roadmap. Full capture/render depends on your local toolchain; offline CI checks sources and selected local utilities, not the complete production stack. See workflow and usage.

Product walkthrough: a local Acme fixture, real browser capture, narration/music, and EN/ES subtitles. Source and commands · Watch with sound.

Acme product walkthrough

Composed CLI/browser film: a multi-environment example whose terminal scenes are authored HTML reenactments. Source and explanation · Watch with sound.

Composed terminal and browser example

The longer examples remain intact. Generated MP4/audio files are ignored; rebuild them from the included sources. See usage prompts for silent, narrated, avatar, voice-swap, and multi-environment requests.

Each video uses projects/<slug>/brief.yaml, brand tokens under brands/<name>/, and generated output under its project folder. Optional private settings live outside Git, selected by DEVREL_VIDEO_ENV.

Setting Purpose
brief.yaml Product URL, audience, message, beats, brand, duration, and optional layers
ORPHEUS_URL / ORPHEUS_KEY Your optional narration endpoint and token
AVATAR_HOST / LS_DIR Your optional SSH GPU host and LatentSync checkout
BGM_STYLE Local background-music style; default corporate

See configuration, the actual brief template, and security/privacy guidance. Keep private footage and credentials outside this public repository.

Brief → explore → brand → capture → compose → optional layers → render/QA → thumbnail. The skill coordinates existing tools rather than replacing them. Detailed production steps are in the workflow guide and production notes.

.claude/skills/devrel-video/  Skill instructions, scripts, templates, production notes
brands/                     Brand tokens and logos
projects/                   Curated public examples; private projects stay ignored
examples/                   Real harness recording, public input, generated brief
assets/                     Shared banner, social card, and short terminal GIF
docs/                       Onboarding, usage, configuration, troubleshooting

Use Issues for questions, bugs, and ideas. Read CONTRIBUTING.md and CODE_OF_CONDUCT.md; report vulnerabilities privately through SECURITY.md.

Install requirements-dev.txt in a virtual environment and run python3 scripts/check.py. See CHANGELOG.md for changes. Live recordings, service calls, GPU jobs, and full renders are manual validation steps.

MIT for original studio content. Embedded banner fonts retain their OFL notices; external tools/models/media retain their own terms. See licensing.


Built by Aitor Argomaniz · LinkedIn · More open source

联系我们 contact @ memedata.com