Claude Code vs Codex vs claude.ai Web Installation: Deploying the Watch Skill

Claude Code installs via built-in marketplace commands, Codex uses the npx skills CLI, and claude.ai Web requires manual upload of a release archive, yet all three methods deploy identical files from the skills/watch/ directory.

The bradautomates/claude-video repository distributes a video processing skill that operates across multiple AI hosting environments. While the underlying code in skills/watch/ remains constant, the installation mechanism varies significantly between Claude Code, Codex, and the claude.ai web interface. Understanding these platform-specific deployment paths ensures proper setup of dependencies like ffmpeg and yt-dlp.

Claude Code Marketplace Installation

Claude Code distributes skills through its built-in marketplace system, utilizing the manifest defined in .claude-plugin/plugin.json. The IDE handles skill discovery, download, and extraction automatically when you execute the registration commands.

/plugin marketplace add bradautomates/claude-video
/plugin install watch@claude-video

The marketplace pulls the pre-packed watch.skill bundle from the repository's release assets and extracts the entire skills/watch/ directory into Claude Code's internal plugin folder. Updates follow the same pattern using /plugin update watch@claude-video.

Codex and Agent Skills CLI Installation

For Codex, Cursor, Copilot, and other Agent Skills hosts, installation occurs through the generic npx skills CLI tool. This method reads the contract from skills/watch/SKILL.md and the manifest from .codex-plugin/plugin.json to determine installation parameters.

# Global install available to all projects

npx skills add bradautomates/claude-video -g

# Preview available skills without installing

npx skills add bradautomates/claude-video -l

# Project-local installation only

npx skills add bradautomates/claude-video

The CLI copies the skills/watch/ folder directly into the host-specific skills directory, typically ~/.codex/skills/ for global installs or a local project directory. This file-tree approach requires no centralized marketplace. Update the skill later using npx skills update watch -g.

claude.ai Web Manual Upload

The claude.ai web interface cannot execute marketplace commands or CLI tools, requiring manual upload of a packaged archive. The repository's skills/watch/scripts/build-skill.sh generates the dist/watch.skill zip file for this purpose.

1. Download the latest watch.skill archive from the repository's releases page.
2. Navigate to Settings → Capabilities → Skills in the claude.ai web UI.
3. Click + and drop the watch.skill file to extract and register the skill.

Because the web UI sandbox cannot invoke system commands by default, you must enable "Code execution and file creation" in settings to allow the skill to execute ffmpeg and yt-dlp at runtime.

Common Runtime Layout

Regardless of the installation method, all three platforms ultimately produce the same runtime structure. The host reads skills/watch/SKILL.md as the contract to discover entry points, while the scripts handle video downloading, frame extraction, and transcription.

skills/watch/
├── SKILL.md               ← contract read by all hosts
└── scripts/
    ├── watch.py           ← main orchestrator
    ├── download.py
    ├── frames.py
    ├── transcribe.py
    ├── whisper.py
    ├── config.py
    ├── setup.py
    └── build-skill.sh     ← packs the archive for web upload

Key Configuration Files

• .claude-plugin/plugin.json: Manifest used by Claude Code's marketplace system to register the skill.
• .codex-plugin/plugin.json: Configuration file read by the npx skills CLI for Codex-family hosts.
• skills/watch/SKILL.md: The universal contract that defines entry points and capabilities for all hosting platforms.
• skills/watch/scripts/build-skill.sh: Build script that generates the dist/watch.skill archive required for claude.ai Web installation.

Summary

• Claude Code uses /plugin commands and a marketplace system to auto-extract skills into the IDE's plugin directory.
• Codex and Agent Skills hosts rely on npx skills to copy the skills/watch/ folder directly into ~/.codex/skills/ or local project directories.
• claude.ai Web requires downloading watch.skill from releases and manually uploading via the Settings UI, with explicit permission grants for code execution.
• All platforms ultimately execute the same scripts from skills/watch/scripts/, including watch.py as the main orchestrator.

Frequently Asked Questions

Can I install the same skill on all three platforms simultaneously?

Yes. The bradautomates/claude-video repository maintains identical code across all distribution methods. You can install via Claude Code's marketplace for local IDE work, use npx skills for Codex projects, and upload the archive to claude.ai Web without conflicts, as each maintains separate runtime environments.

Why does claude.ai Web require enabling "Code execution and file creation"?

The web interface runs in a sandboxed environment that blocks system command execution by default. Since the watch skill invokes external binaries like ffmpeg for video processing and yt-dlp for downloads, you must explicitly grant code execution permissions in Settings → Capabilities for these subprocess calls to function.

Where is the skill installed when using the Codex CLI globally?

When using npx skills add bradautomates/claude-video -g, the CLI copies the entire skills/watch/ directory into the host's global skills folder, typically located at ~/.codex/skills/ on Unix systems. Local installs place the folder in a .skills/ directory within your current project root instead.

How do I update the skill after installation?

For Claude Code, run /plugin update watch@claude-video. For Codex and Agent Skills hosts, execute npx skills update watch -g for global updates or npx skills update watch for local project updates. For claude.ai Web, you must manually download the new release archive and re-upload it through the Skills settings panel to replace the existing installation.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}