Video Transcriber

After output (DETECT): Invoke /fact-checker to verify names, institutions, and legislation references. For transcripts, verify all person names against ~/.claude/skills/bangsamoro/references/barmm-officials-2025-2026.md.

Before publishing (CONFIRM): Present verification results to the user.

Prerequisites

The transcribe script handles dependency checking and will tell you what's missing. The required tools:

# Primary (required)
pip3 install youtube-transcript-api

# Fallback (optional — only needed for videos without YouTube captions)
brew install yt-dlp          # or: pip3 install yt-dlp
pip3 install openai-whisper  # requires ffmpeg: brew install ffmpeg

Usage

Run the transcription script from this skill's directory:

SKILL_DIR="/Users/saidamenmambayao/.claude/skills/youtube"
"$SKILL_DIR/venv/bin/python3" "$SKILL_DIR/scripts/transcribe.py" [OPTIONS] URL [URL ...]

Options

Examples

# Basic transcription
python3 scripts/transcribe.py "https://youtube.com/watch?v=abc123"

# Filipino language transcript
python3 scripts/transcribe.py --lang fil "https://youtube.com/watch?v=abc123"

# Save to file
python3 scripts/transcribe.py --output transcript.md "https://youtube.com/watch?v=abc123"

# With Whisper fallback enabled
python3 scripts/transcribe.py --fallback "https://youtube.com/watch?v=abc123"

# Batch — multiple videos
python3 scripts/transcribe.py URL1 URL2 URL3

# Batch — save each to separate files
python3 scripts/transcribe.py --output transcripts/ URL1 URL2 URL3

Output Format

The final document has THREE sections:

1. Organized Notes (top) — A detailed, AI-generated summary organized by topic with bold section headers and clean prose. NO timestamps in this section. Claude generates this AFTER transcription by reading the Transcript section.

2. What This Means for Your Work (middle) — A personalized evaluation of how the video's knowledge applies to the user's actual work. Before writing this section, read ~/Vault/about-saidamen.md for the user's full profile — roles, projects, competencies, and what drives the work. Reference specific projects, roles, and initiatives from the profile (not generic advice). Analytical and actionable — not a summary, but an assessment of what changes, validates, or challenges the user's current approach. Include concrete next steps tied to specific projects or workflows the user actually has.

   Required subsection — "How This Can Improve Your Claude Skills and Workflows": After the main analysis, add an ### How This Can Improve Your Claude Skills and Workflows subsection. This evaluates how the video's ideas could enhance, refine, or challenge the user's existing Claude Code skill system and AI-augmented workflows. Before writing, consult ~/Vault/Claude-Skills/index.md for the full 124-skill library and ~/Vault/about-saidamen.md for the user's Claude Code setup (129 skills, Obsidian vault memory, NotebookLM integration, parallel agents, frontend-first development, fact-checking pipeline).

   Reference specific skills and projects by name. The subsection should cover any combination of:

   • Skill improvements — which existing skills (e.g., /research-pipeline, /expert-builder, /skill-optimizer, /fact-checker, /bangsamoro, /bill-drafter) could benefit from ideas in the video
   • New skill opportunities — does the video suggest a skill the user doesn't have yet? Check ~/Vault/Claude-Skills/index.md to verify it doesn't already exist before recommending
   • Workflow changes — how could the video's patterns change the user's development workflow (frontend-first dev, parallel agents, Obsidian vault sync, NotebookLM pipeline, skills-bucket)
   • Project-specific applications — which of the 9+ active projects (e-Bangsamoro, MoroMarket, Parliamentarian, OBCMS, Tarbiyyah-MS, BangsamoroHR, IPP, SBP, MoroTech) would benefit from implementing the video's ideas
   • Cross-pollination — connections between the video's domain and the user's Bangsamoro governance, cooperative development, or legislative capacity work

   Keep it analytical and specific — name the skills, name the projects, describe the concrete change. No generic "this could improve your workflow" statements.

# Video Title

**Channel**: Channel Name
**Duration**: 12:34
**Language**: English (auto-generated)
**URL**: https://youtube.com/watch?v=abc123

---

## Organized Notes

**Introduction**

- Key idea expressed as a **concise bullet** with bold keywords
- Another key point with **specific examples** preserved

**Key Concept Name**

- Main idea of this concept in a **scannable bullet**
- Steps or sub-points as a **numbered list**:
  1. First step
  2. Second step
- Takeaway: **bold the key insight**

**Another Topic**

- More key ideas as **bullets, not paragraphs**

---

## What This Means for Your Work

Personalized evaluation of how this video's knowledge applies to the user's
specific work context. Analytical, not summary. Includes concrete next steps.

### How This Can Improve Your Claude Skills and Workflows

Specific analysis of how the video's ideas map to existing skills (by name),
active projects, and AI-augmented workflows. References skills from
~/Vault/Claude-Skills/index.md and projects from ~/Vault/about-saidamen.md.

---

## Transcript

[00:00] Complete sentences grouped into idea-sized paragraphs.
Multiple caption fragments are merged into coherent blocks.

[00:13] Another complete idea with its own timestamp. Each block
represents a logical thought, not a raw caption segment.

Workflow

When the user provides a YouTube URL:

0. Duplicate check: Before transcribing, extract the video ID from the URL and grep docs/video-transcripts/index.md for it. The index is a markdown file with tables organized by topic subfolder, where each row has | VIDEO_ID | Title | Filename |.

   grep "VIDEO_ID" docs/video-transcripts/index.md
   

   • If found: tell the user "This video has already been transcribed" and show the existing file path. Ask if they want to re-transcribe (overwrite) or skip.
   • If not found: proceed to step 1.
   • Extract video ID from URL: youtube.com/watch?v=VIDEO_ID or youtu.be/VIDEO_ID

1. Run the transcribe script with the user's URL and any preferences (language, output path). The script produces the metadata header + placeholder Organized Notes + idea-grouped Transcript.

2. Generate Organized Notes: After the script saves the file, read the Transcript section and generate the Organized Notes to replace the <!-- PLACEHOLDER --> comment. Follow these rules:

   • Create meaningful bold section headers that group content by topic
   • Use bullet points and numbered lists — NOT dense paragraphs
   • Bold key words and key phrases within bullets for scannability
   • NO timestamps in this section
   • Capture every key idea, example, and argument — but keep each bullet concise
   • Preserve specific names, tools, frameworks, and examples mentioned
   • First mention rule: When a person is first mentioned, provide their complete name and full position/title (e.g., "Interim Chief Minister Abdulraof A. Macacua", "CEO John Smith of Acme Corp"). Subsequent mentions can use shortened forms. Research correct names/titles if auto-captions garble them.
   • Organize by topic, not chronologically
   • Language cleanup: Remove garbled characters from non-expected scripts (Thai, Bengali, Hindi, etc.) — these are auto-caption noise. Only English, Filipino, and Arabic are expected.
   • Arabic translation: When Arabic text appears (prayers, Quranic verses, formal greetings), preserve the Arabic and provide an English translation in brackets immediately after (e.g., [Translation: O Allah, guide our leaders...])
   • : Place quotes (NOT as a separate bottom section). Add up to 5 quotes per speaker labeled . Each quote should be standalone and complete (1-5 sentences), meeting 3+ criteria: standalone clarity, emotional punch, policy significance, quotable length, attributable, shareable on social media. Format: Clean up filler words but preserve the speaker's voice.

/guide-drafter {topic-folder} {new-transcript-path}

The guide-drafter reads the existing GUIDE, determines what the new transcript adds (new patterns or new evidence for existing patterns), updates with proper Chicago/Turabian footnote citations, and saves to both repo and vault. If no GUIDE exists yet for this topic group, guide-drafter runs in CREATE mode to generate one from scratch.

Gotchas

• Filipino Muslim names from auto-captions are ALWAYS garbled — verify every person name against ~/.claude/skills/bangsamoro/references/barmm-officials-2025-2026.md
• Thai, Bengali, Hindi, and other non-expected script characters are auto-caption noise — delete them; only English, Filipino, and Arabic are expected
• Arabic prayers and Quranic verses should be preserved in Arabic with English translation in brackets
• Channel names are not always the speaker's name — extract the actual host/creator from transcript content for citation metadata
• Whisper model "base" is the default fallback — for long videos (1h+), the primary YouTube caption API method is instant while Whisper takes minutes
• Auto-generated captions frequently merge words or split them incorrectly — always proofread proper nouns, legislation references, and technical terms

Error Handling

Tips

• Most YouTube videos have auto-generated English captions — the primary method works almost always
• For non-English videos, try --lang with the video's language code first
• The base Whisper model is a good balance of speed and accuracy for the fallback
• For long videos (1h+), the primary method is instant; Whisper fallback will take several minutes
• Batch mode with --output transcripts/ creates one .md file per video, named after the video title
• YouTube chapters (when set by the creator) can inform section header choices in Organized Notes

Transcript (bottom) — Verbatim [MM:SS] timestamped idea-grouped paragraphs, produced by the script. Each timestamp marks a complete idea (a sentence or paragraph), not a raw caption fragment.

*Quotable Quotes:*
> "Standalone complete quote here." [MM:SS]