Yt Summarize

Step 1: Parse Arguments

Extract from $ARGUMENTS:

• URL (required): YouTube video URL, video ID, or playlist URL — first positional argument
• --mode: Summary mode — kurz (brief), standard (default), learn
• --lang: Subtitle languages, comma-separated (default: de,en)
• --engine: Transcription engine — auto (default), whisper, whisper-api, gpt4o-transcribe
• --whisper: Shortcut for --engine whisper
• --transcript-only: Output transcript without summarization
• --summary-lang: Language for the summary (default: Deutsch)
• --model: Claude model for summarization (default: claude-sonnet-4-6)
• --save-learnings: Save learn mode output to knowledge base

If no URL is provided, ask the user for one.

Summary Modes

Engine Options

Playlist Support

The script auto-detects playlist URLs (containing list= or /playlist?). When a playlist URL is provided:

• All videos are extracted using yt-dlp --flat-playlist
• Each video is processed sequentially (transcribe → summarize)
• Errors on individual videos are logged but don't stop the playlist
• A final summary shows succeeded/failed counts

The script handles everything — just pass the playlist URL as the url argument. All flags (--mode, --engine, --save-learnings, etc.) apply to every video in the playlist.

For the interest feedback loop (Step 5): When processing a playlist, ask for interest feedback once at the end covering all videos, not after each individual video. Summarize the main topics across all videos and let the user rate overall.

Step 2: Transcribe

Run the script to get the transcript. Use $CLAUDE_PLUGIN_ROOT for the script path:

With specific engine:

python3 "${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/skills/yt-summarize}/scripts/yt-summarize.py" "$URL" --engine "$ENGINE" --transcript-only

Auto mode (default — subtitles first, then fallback):

python3 "${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/skills/yt-summarize}/scripts/yt-summarize.py" "$URL" --lang "$LANG" --transcript-only

IMPORTANT: Non-interactive fallback handling

In auto mode, if no subtitles are found and the script cannot prompt interactively (no TTY), it exits with code 2 and prints available engines as JSON to stderr.

When exit code is 2:

1. Parse the JSON from stderr to get available engines
2. Use AskUserQuestion to ask the user which engine to use
3. Re-run with --engine <chosen_engine>

Step 3: Summarize

Run the full pipeline with the chosen mode:

python3 "${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/skills/yt-summarize}/scripts/yt-summarize.py" "$URL" --mode "$MODE" --lang "$LANG" --engine "$ENGINE" --summary-lang "$SUMMARY_LANG" --model "$MODEL"

For learn mode with knowledge base saving:

python3 "${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/skills/yt-summarize}/scripts/yt-summarize.py" "$URL" --mode learn --save-learnings

Step 4: Output

Present the result clearly formatted. If the transcript is very long (>50,000 chars), warn the user about potential token usage before summarizing.

Mode-specific output:

kurz: Display the summary directly.

standard: Display the summary. Due to length (~2000 words), consider using a collapsible section or clear heading structure.

learn: Display the learning summary. If --save-learnings was used, inform the user where files were saved. If skill matches were found, highlight them.

Step 5: Interest Feedback Loop

After presenting ANY summary (all modes), always do this:

1. Use AskUserQuestion to ask:

   "Wie interessant war dieses Video für dich?" Options:

   • Kein Interesse — "Thema ist nicht relevant für mich"
   • Wenig Interesse — "Etwas relevant, aber keine Priorität"
   • Hohes Interesse — "Sehr relevant, mehr davon!"

2. Based on the response, update INTERESTS.md:

   • Read the current INTERESTS.md from the repo (find it via $CLAUDE_CONFIG_REPO/INTERESTS.md or search for it)
   • Extract the main topics/tags from the video summary
   • Hohes Interesse: Add or reinforce topics under ## Topics of Interest with today's date
     • If topic already exists, add today's date as reinforced: YYYY-MM-DD
     • If new topic, add as (added: YYYY-MM-DD)
   • Kein Interesse: Add topics under ## Not Interested with (since: YYYY-MM-DD)
   • Wenig Interesse: No changes (neutral signal)

3. If INTERESTS.md was changed:

   • Write the updated file
   • Commit with message: chore: update INTERESTS.md from yt-summarize feedback
   • Tell the user: "Dein Interessenprofil wurde aktualisiert."

Error Handling

• No yt-dlp: Tell user to install with uv pip install yt-dlp
• No whisper: Tell user to install with uv pip install openai-whisper
• No openai: Tell user to install with uv pip install openai
• No ANTHROPIC_API_KEY: Remind user to set it (only needed for summarization)
• No OPENAI_API_KEY: Remind user to set it (only needed for API engines)
• No engines available: Suggest installing either whisper or openai package
• Exit code 2: Engine selection needed — use AskUserQuestion (see Step 2)
• Timeout: Report which command timed out and suggest retry
• Private/unavailable video: Report the yt-dlp error clearly

Self-Improvement Protocol

When you notice something that would improve this skill:

1. Append the insight to IMPROVEMENTS.md with date and context
2. Do NOT modify SKILL.md directly
3. Tell the user: "I logged a potential improvement for yt-summarize"