buda-youtube-to-script

这段内容是审核时保存的快照。外部来源才是完整且最新的版本。

---
name: buda-youtube-to-script
description: >-
  Use this skill whenever the user provides a YouTube or other video URL and
  asks to create, generate, rewrite, repurpose, summarize into, or get a new
  talking script, voiceover script, or short-form video script. Trigger
  strongly when this skill is selected and the user sends only a video URL,
  or on phrases like "get new talking script", "rewrite this video as
  a script", "turn this YouTube video into a voiceover", or when the user
  gives a video URL plus an angle or prompt. This skill is fully
  self-contained — it downloads the source video and audio with
  the bundled yt-dlp helper, transcribes the audio with Whisper, reads the
  bundled assets/talking-script-template.md, then produces a new
  template-based talking script. No other skill needs to be installed.
---

# YouTube to Talking Script

Turn a video URL into a new talking script by downloading the source video and audio, transcribing it, and rewriting the content into the user's requested angle.

This skill is self-contained. It does NOT depend on any separate downloader skill. The yt-dlp download logic ships with this skill at `scripts/download-audio.sh`, and the output template ships at `assets/talking-script-template.md`.

Important default: for Buda demos and script-generation tasks, keep the source video as a user-facing artifact together with the MP3, transcript, and final script.

## Default behavior

Unless the user says otherwise:

- Output language: English
- Script style: short-form talking script suitable for Red Note / TikTok / YouTube Shorts
- Target length: 60–90 seconds
- Rewrite mode: creative repurposing, not a literal translation or transcript summary
- Preserve the source video's core idea, but make the final script natural for a human to speak
- Include a strong 0–3 second hook
- Save all artifacts under `/agent/outputs/talking-scripts/<safe-title-or-id>-<YYYYMMDD-HHMMSS>/`
- By default keep the downloaded source video, MP3 audio, transcript, and final script

## Inputs to extract from the user

Required:

- Video URL

Optional:

- Angle / prompt, for example: `get new talking script`, `for entrepreneurs`, `make it 60 seconds`, `Red Note style`, `focus on how AI tools boost efficiency`
- Target platform
- Target audience
- Target duration
- Output language
- Whether to include timestamps, subtitles, or title ideas

If the user gives only a URL, proceed with the defaults. Do not ask unnecessary clarification questions.

## Workflow

### 1. Prepare a task folder

Create a durable output directory, not scratch files in `/agent` root.

```bash
mkdir -p /agent/outputs/talking-scripts
WORKDIR="/agent/outputs/talking-scripts/$(date +%Y%m%d-%H%M%S)"
mkdir -p "$WORKDIR"
```

If possible, use `yt-dlp --print title` or `yt-dlp --print id` to create a more descriptive folder name. Sanitize names so they are safe as file paths.

### 2. Verify dependencies

```bash
which ffmpeg || echo "ffmpeg missing — install with: brew install ffmpeg"
```

You normally do NOT need to install yt-dlp manually. The bundled download script resolves `yt-dlp` in this order:

1. `yt-dlp` on PATH
2. `/agent/.local/bin/yt-dlp`
3. If still missing, it auto-installs into a dedicated venv at `/agent/.local/venvs/yt-dlp` and symlinks `/agent/.local/bin/yt-dlp` — this persists across sandbox restarts.

Auto-install needs `python3` (or `python`) available, which the Buda sandbox provides by default. `ffmpeg` is required for MP3 extraction and is also preinstalled.

### 3. Download source video and audio with the bundled helper

Keep both source video and MP3 audio as user-facing artifacts. Use the bundled script. Resolve its path against this skill directory:

```text
/space/.agents/skills/buda-youtube-to-script/scripts/download-audio.sh
```

```bash
bash /space/.agents/skills/buda-youtube-to-script/scripts/download-audio.sh \
  -o "$WORKDIR" --video \
  "VIDEO_URL"
```

The helper:

- Saves the source video to `$WORKDIR/video/`