Video to Article Conversion Workflow

This skill automates the conversion of lecture/presentation videos into various text formats including transcripts, outlines, and article drafts.

Workflow Overview

1. Setup Folder - Create properly named folder if needed
2. Validate Input Metadata - Check for README.md with required frontmatter
3. Extract Audio - Convert video to MP3 using ffmpeg
4. Transcribe - Generate text transcript using ElevenLabs API
5. Process Text - Create progressive text refinements

Step 0: Folder Setup (if starting from scratch)

If user is starting a new talk/lecture conversion and no folder exists yet:

Create folder using format: YYYY Title of Talk

After creating folder, navigate into it and proceed with workflow.

Step 1: Validate Metadata

Look for README.md in the current directory with these frontmatter fields:

• title - Lecture/presentation title
• speaker - Speaker name
• date - Presentation date (YYYY-MM-DD)
• lang - Presentation language (e.g., "english", "czech", "spanish")
• abstract - Brief description
• slides - Link to slides (or file path)
• video - Optional: link to video

If README.md is Missing or Incomplete

Use the AskUserQuestion tool to collect missing information:

- What is the lecture title?
- Who is the speaker?
- What date was this presented?
- What language is the presentation in?
- Brief description or abstract?
- Link to slides or presentation file?

Create or update README.md with collected information.

Step 2: Find or Download Video File

Check for Local Video

Look for video file in current directory:

• Common patterns: *.mp4, *.mov, *.avi, *.mkv
• If multiple found, ask user which to process

If No Local Video Found

Ask user if they have a YouTube URL using AskUserQuestion tool:

• "Do you have a YouTube URL for this talk?"
• If yes, collect the YouTube URL
• If no, ask for local video file path

Download from YouTube (if URL provided)

bash ${CLAUDE_PLUGIN_ROOT}/skills/video-to-article/scripts/download-youtube.sh <youtube-url>

This downloads the video as video.mp4 in medium quality (720p max).

Requirements: User needs yt-dlp installed. If missing, show:

Install yt-dlp:
- pip install yt-dlp
- brew install yt-dlp (macOS)
- sudo apt install yt-dlp (Ubuntu/Debian)

Step 3: Extract Audio

Check if audio.mp3 already exists. If not:

Run the extract-audio.sh script using the plugin root variable:

bash ${CLAUDE_PLUGIN_ROOT}/skills/video-to-article/scripts/extract-audio.sh <video-file>

This creates audio.mp3 in the current directory.

Step 4: Transcribe Audio

Check if generated-transcript.txt already exists. If not:

Map Language to ISO Code

Convert the lang field from README.md to the appropriate ISO 639-1 language code for ElevenLabs API. Use your knowledge of language codes to map language names to their two-letter codes (e.g., "english" → "en", "czech" → "cs", "spanish" → "es").

Run Transcription

bash ${CLAUDE_PLUGIN_ROOT}/skills/video-to-article/scripts/transcribe-audio.sh <language-code>

Step 5: Generate Text Outputs

Execution Protocol

Before generating content, complete these steps in order:

1. List all required outputs - Read and acknowledge the complete list below
2. Check which files already exist - Use file existence checks for each output
3. Read all existing source files - Load transcript, outline, key-ideas as context
4. Confirm which outputs need generation - Explicitly state which files are missing
5. Only then begin generation - Generate missing files one by one

Required Outputs

Generate all of these files:

1. generated-transcript-cleaned.md - Cleaned transcript
2. generated-transcript-readable.md - Readable transcript
3. generated-transcript-outline.md - Content outline
4. generated-key-ideas.md - Key takeaways
5. generated-article-formal.md - Explanatory blog post (long-form, narrative)
6. generated-article-direct.md - Scannable technical brief (concise, bullets)

Note: Outputs 5 and 6 are two different article variants with distinct styles and audiences. Generate both.

Workflow:

• Check file existence for all files above
• Automatically generate any missing files without asking user
• Skip files that already exist (inform user which were skipped)
• Don't declare completion unless all files exist
• Don't ask user if they want to generate missing files - just generate them
• Generate both article variants (formal and direct) - they serve different purposes

Idempotent Generation Strategy

This process is idempotent - running the skill multiple times will only generate missing files, never overwrite existing ones:

1. Check each output file before generating - Use file existence checks
2. Skip files that already exist - Inform user which files are being skipped
3. Generate only missing files - Don't overwrite existing content
4. Use existing files as context - When generating later outputs (like key-ideas), read all available previous outputs for context

Example: If only generated-key-ideas.md is missing:

• Read existing: generated-transcript-cleaned.md, generated-transcript-readable.md, generated-transcript-outline.md
• Generate: generated-key-ideas.md using the existing files as context
• Skip: All other outputs

This allows users to:

• Resume interrupted workflows
• Generate only specific missing outputs
• Re-run safely without data loss

Output 1: Cleaned Transcript

File: generated-transcript-cleaned.md Input: generated-transcript.txt Check: Skip if file exists

Convert the transcript into a more readable form by dividing it into paragraphs and adding headings.

Instructions:

• Do not change the content, except for obvious typos
• Do not change the tone of voice - if the speaker uses informal language, keep it
• Remove filler words like "Eh", "Ehm", "Um", "Uh"
• Remove comments related to the transcription itself, or rewrite them elegantly if possible (e.g., "(laughter)" or "(2 second pause)")
• Respond in {lang} language - maintain the same language as the original transcript

The cleaned transcript should preserve the speaker's authentic voice while being easier to read and follow.

Output 2: Readable Transcript

File: generated-transcript-readable.md Inputs: generated-transcript.txt and generated-transcript-cleaned.md Check: Skip if file exists

You have two inputs:

1. The first text contains the direct transcript of a lecture
2. The second text contains a better structured version of this transcript

The slides are available here: {slides}

Create a more readable text from this transcript:

Instructions:

• Preserve the style, but optimize for reading comprehension
• Fix typos, repetitions, and improve stylistic clarity
• Shorten where it doesn't affect the message
• Make it flow naturally for readers while keeping the speaker's voice
• Respond in {lang} language - the output must be in the same language as the input

The goal is to make the content accessible to readers while maintaining the essence of the spoken presentation.

Output 3: Outline

File: generated-transcript-outline.md Inputs: Cleaned and readable transcripts Check: Skip if file exists

From the provided transcript inputs, create an outline of what was discussed.

Instructions:

• List the main ideas and messages with very brief content
• Use bullet points for clarity
• Focus on key takeaways and core concepts
• Keep descriptions concise - just enough to understand the point
• Respond in {lang} language - the outline must be in the same language as the source material

This outline should serve as a quick reference guide to the presentation's structure and main points.

Output 4: Key Ideas

File: generated-key-ideas.md Inputs: Cleaned and readable transcripts (or all available generated files if running incrementally) Check: Skip if file exists

From the provided transcript inputs, extract the key ideas, tips, and main concepts.

Instructions:

• Focus on interesting insights, actionable tips, and core concepts
• Unlike the outline, don't follow chronological order - prioritize importance
• Exclude generic or procedural content
• Each idea should have a brief description explaining why it matters
• Use bullet points or numbered list
• Respond in {lang} language - the output must be in the same language as the source material

This should capture the most valuable takeaways someone would want to remember from the talk.

Output 5: Explanatory Blog Post

File: generated-article-formal.md Inputs: All previous outputs (or all available generated files if running incrementally) Check: Skip if file exists

Create a traditional blog post with complete explanations and smooth narrative flow.

Target audience: General readers discovering the topic, browsing casually Reading context: Website visitors who have time and want full understanding

Instructions:

• Write in complete, flowing sentences with careful transitions
• Provide context and background before diving into details
• Use storytelling elements to maintain engagement
• Example opening: "V průběhu posledního roku jsme zaznamenali značné změny..." (Czech) or equivalent
• Explain why concepts matter, not just what they are
• Guide readers through ideas with clear signposting
• Respond in {lang} language - the article must be in the same language as the source material

Style characteristics:

• Narrative structure with introduction, development, conclusion
• Welcoming tone that assumes less prior knowledge
• Longer paragraphs with descriptive language
• Educational approach that builds understanding step-by-step

Output 6: Scannable Technical Brief

File: generated-article-direct.md Inputs: All previous outputs (or all available generated files if running incrementally) Check: Skip if file exists

Create a condensed, high-signal article optimized for rapid scanning and information extraction.

Target audience: Technical readers or busy professionals who skim content Reading context: Newsletter, Slack/Discord shares, or quick reference

Instructions:

• Lead with short, declarative statements - skip warm-up sentences
• Frontload key information in each paragraph
• Heavy use of bullets, numbered lists, and visual breaks
• Example opening: "AI modely: radikální skok za rok." (Czech) or equivalent
• Cut all filler words, hedging, and obvious statements
• Dense information - every sentence earns its place
• Respond in {lang} language - the article must be in the same language as the source material

Style characteristics:

• Telegraphic style - fragments are encouraged when clearer
• Bullets and lists dominate over prose paragraphs
• Assumes reader familiarity with domain
• Optimized for 2-minute skim, not 10-minute read

Summary

After completion, inform the user which files were:

• Skipped (already existed)
• Generated (newly created)

Example output:

Workflow complete!

Skipped (already exists):
- audio.mp3
- generated-transcript.txt
- generated-transcript-cleaned.md
- generated-transcript-readable.md
- generated-transcript-outline.md

Generated:
- generated-key-ideas.md
- generated-article-formal.md
- generated-article-direct.md

All outputs in <language> language.

If everything was generated fresh, simply list all outputs under "Generated".

Error Handling

• Missing ffmpeg: Prompt user to install ffmpeg
• Missing API key: Prompt user to set ELEVENLABS_API_KEY
• Transcription errors: Show error message and suggest checking API key/quota
• Script not found: Use absolute paths to skill scripts directory