Stream Transcript Processor

Stream Transcript Processor

Process @techfren stream transcripts into actionable content assets: clips, script notes, X posts, YouTube metadata, and analytics.

This skill orchestrates existing skills and adds new capabilities:

• Uses clipper for clip extraction
• Uses x-post-extractor patterns for X posts
• Uses media-fetcher for screenshots and media
• NEW: Fetch transcripts from URLs (auto-captions)
• NEW: Generate script notes for short-form recording
• NEW: YouTube metadata (title, description with timestamps, SEO tags)
• NEW: Content analytics and insights

When to Use This Skill

Activate when user:

• Provides a Twitch VOD or YouTube stream URL
• Pastes a stream transcript
• Asks to "find clips" or "find highlights" from a stream
• Wants "script notes" for short-form content
• Asks for X post drafts from stream content
• Needs YouTube title, description, or tags
• Requests stream analytics or insights

Quick Commands

Execution Workflow

Full Pipeline (URL Provided)

When user provides a stream URL:

Phase 1: Fetch Transcript

python .claude/skills/stream-transcript-processor/scripts/fetch_transcript.py "<URL>" transcript.txt

This creates:

• transcript.txt - Timestamped text file
• transcript.json - Structured JSON for processing

Phase 2: Parse for Clipper Compatibility

python .claude/skills/clipper/scripts/parse_transcription.py transcript.json > parsed.json

Phase 3: Find Clips (Use Clipper Skill)

Follow the clipper skill workflow:

1. Analyze parsed.json in windows
2. Identify segments using clipper categories + voice patterns
3. Create segments.json
4. Report clip candidates to user

Phase 4: Generate Content Assets

For each high-priority clip:

Script Notes:

python .claude/skills/stream-transcript-processor/scripts/generate_script_notes.py parsed.json <start_ts> <end_ts>

X Post Drafts:

python .claude/skills/stream-transcript-processor/scripts/draft_x_posts.py parsed.json <timestamp> --style discovery

Phase 5: Fetch Media (Optional)

Use media-fetcher for screenshots of tools mentioned:

python .claude/skills/media-fetcher/scripts/capture_screenshot.py "<url>" --output media/

Transcript Already Exists

If parsed.json or out.json already exists:

1. Skip fetch_transcript step
2. Use existing files for analysis
3. Proceed with clipper workflow
4. Add script notes and X posts as needed

Script Notes Only

When user asks for script notes for a specific segment:

python .claude/skills/stream-transcript-processor/scripts/generate_script_notes.py \
    <transcript_file> <start_timestamp> <end_timestamp> \
    --output script_notes.md

Output format (sentence starters for recording):

## HOOK
- [Bold claim or result]...
- This is [Tool] and it...

## VALUE STACK
- Free, open source...
- Runs locally...

## DEMO
- Let's try it...
- [React naturally]

## CLOSE
- [Quick recap]...
- Try it out...

X Posts Only

When user asks to draft X posts from stream:

Option 1: Use x-post-extractor skill (recommended for full workflow)

Follow x-post-extractor SKILL.md for comprehensive extraction.

Option 2: Quick draft from timestamp

python .claude/skills/stream-transcript-processor/scripts/draft_x_posts.py \
    <transcript_file> <timestamp> --style <style>

Styles: discovery, result, comparison, tutorial

YouTube Metadata Generation

Generate SEO-optimized YouTube title, description with timestamps, and tags:

python .claude/skills/stream-transcript-processor/scripts/generate_youtube_metadata.py \
    <transcript_file> [--title "Custom Title"] [--output youtube_metadata.md]

Output includes:

• Engaging title with hook patterns
• Description with auto-generated timestamps
• Tool links for all mentioned products
• Social links section
• SEO-optimized tags (up to 30)

Example output:

## Title

Claude Haiku Benchmark: New Record! | Live Testing

## Description

🎯 Claude Haiku Benchmark: New Record! | Live Testing

In this stream, I test and demo various AI tools and share my findings live. Watch to see real-time benchmarks, comparisons, and honest reactions.

⏱️ TIMESTAMPS 0:00 - Intro 6:12 - Setting up Playwright MCP 11:05 - Claude Code $200 plan discussion 47:35 - Testing Playrider MCP 1:01:31 - Playrider speed results 1:04:29 - Haiku 43 second record!

🔗 TOOLS MENTIONED • Claude: https://claude.ai • Playwright: https://playwright.dev • Augment: https://augmentcode.com

📱 CONNECT • Twitter/X: https://x.com/techfren • GitHub: https://github.com/aj47 • Website: https://techfren.net

#AI #Coding #Programming #Claude #Benchmark #TechStream

## Tags

claude, haiku, benchmark, ai coding, playwright mcp, live coding, ai tools, programming, developer tools, automation

See youtube-description-template.md for full template and SEO guidelines.

Analytics Report

When user asks to "analyze" a stream, generate insights:

Report Template

# Stream Analytics: [Title/Date]

## Overview
- Duration: X hours
- Total transcript entries: N
- High-energy moments detected: N

## Topics Covered
| Topic | Mentions | Peak Timestamp |
|-------|----------|----------------|
| Cursor | 15 | 01:23:45 |
| Claude | 12 | 02:15:30 |
| ...

## Energy Peaks
Timestamps with highest concentration of reaction words:
1. [01:23:45] "Dude, this is insane..." (Score: 12)
2. [02:15:30] "Whoa, that actually worked..." (Score: 10)
3. ...

## Clip Potential
- High priority clips: N
- Medium priority clips: N
- Estimated short-form content: N minutes

## Value Props Mentioned
- Free: X times
- Open source: X times
- Local/offline: X times

## Recommended Extractions
1. [Timestamp] - [Brief description] - **Clip + X Post**
2. [Timestamp] - [Brief description] - **Script Notes**
3. ...

Integration with Existing Skills

Using Clipper

For full clip extraction, delegate to clipper:

# Claude internally:
# 1. Ensure parsed.json exists
# 2. Follow clipper SKILL.md workflow
# 3. Create segments.json
# 4. Extract clips

The clipper skill handles:

• Window-based analysis
• Segment identification
• FFmpeg extraction
• Cleanup processing

Using X-Post-Extractor

For comprehensive X post extraction:

# Claude internally:
# 1. Ensure parsed.json exists
# 2. Follow x-post-extractor SKILL.md workflow
# 3. Create x_posts_output/ folder
# 4. Include media capture

The x-post-extractor skill handles:

• Trigger phrase detection
• Scoring and prioritization
• Post/thread formatting
• Media fetching

Using Media-Fetcher

For screenshots and supplementary media:

# Screenshots of tools/websites
python .claude/skills/media-fetcher/scripts/capture_screenshot.py "<url>"

# Fetch logos
python .claude/skills/media-fetcher/scripts/fetch_logos.py "<domain>"

# Extract video clips
ffmpeg -ss <start> -i "<video>" -t <duration> clip.mp4

Voice Pattern Detection

When analyzing transcripts, look for these signals (see voice-patterns.md):

High-Value Signals (Score +4)

• Discovery phrases: "just discovered", "trending", "check this out"
• Comparison signals: "better than", "beats", "vs"
• Resolution markers: "finally", "figured it out", "turns out"

Medium-Value Signals (Score +2-3)

• High-energy reactions: "dude", "whoa", "sick", "boom"
• Value props: "free", "open source", "local"
• Demo phrases: "let's see", "let's try", "here we go"
• Result moments: "it worked", "there we go", "boom"

Tool Detection

TOOLS = [
    'cursor', 'claude', 'gpt', 'copilot', 'windsurf', 'augment',
    'ollama', 'llama', 'mistral', 'groq', 'anthropic', 'openai'
]

Clip Criteria Summary

From clip-criteria.md:

What Makes a Good Clip

• Discovery moments with specific tools/results
• Authentic reactions that feel genuine
• Value props stacked together (free + open source + local)
• Result moments showing proof something works
• Tool demos with live testing

Optimal Lengths

• TikTok/Reels/Shorts: 30-60 seconds
• Longer tutorials: 2-3 minutes
• X video: 30-45 seconds

Skip These

• Slow starts ("so...", "hey what's up...")
• Incomplete thoughts
• Negative content without resolution
• Off-topic tangents

Output Folder Structure

When running full pipeline:

./
├── transcript.txt          # Fetched auto-captions
├── transcript.json         # Structured transcript
├── parsed.json             # Clipper-compatible format
├── segments.json           # Identified clips
├── youtube_metadata.md     # Title, description, tags
├── clips/                  # Extracted video clips
│   ├── 001_Tool_Demo.mp4
│   ├── 002_Discovery.mp4
│   └── cleaned/
│       └── ...
├── script_notes/           # Recording notes
│   ├── 01_tool_demo.md
│   └── 02_discovery.md
├── x_posts_output/         # X post drafts
│   ├── 01_tool.txt
│   ├── 02_discovery.txt
│   └── README.md
└── media/                  # Screenshots, logos
    ├── cursor.com.png
    └── github_tool.png

Requirements

For Transcript Fetching

• yt-dlp - Download auto-captions (brew install yt-dlp)

For Clip Extraction

• ffmpeg - Video processing (brew install ffmpeg)
• Python 3.8+

For Screenshots (Optional)

• playwright - Browser automation (pip install playwright && playwright install chromium)

Example Usage

Full Pipeline

User: "Process this stream https://youtube.com/watch?v=abc123"

Claude:
1. Fetches transcript → transcript.txt
2. Parses for clipper → parsed.json
3. Analyzes and finds 15 clips
4. Generates script notes for top 5
5. Drafts 3 X posts
6. Captures screenshots for tools mentioned
7. Reports summary with links to all outputs

Quick Script Notes

User: "Generate script notes for the segment at 1:23:45"

Claude:
1. Loads transcript
2. Extracts 60-second window around timestamp
3. Generates HOOK/VALUE/DEMO/CLOSE structure
4. Outputs to terminal or saves to file

Analytics Only

User: "Analyze this stream for clip potential"

Claude:
1. Scans transcript for voice patterns
2. Counts topics, energy peaks, value props
3. Identifies top clip candidates
4. Generates analytics report

Troubleshooting

"No transcript found"

• Video may not have auto-generated captions
• Try different language variants (en-US, en-GB)
• Use full transcription instead (batch-processor skill with tr alias)

"Low clip scores"

• Stream may be low-energy or monotone
• Lower thresholds or look for teaching moments
• Focus on specific tool mentions

"Script notes feel generic"

• Need more specific timestamp with clear tool/demo
• Check if segment has enough context
• May need to expand timestamp window

Supporting Documents

• voice-patterns.md - Detection patterns for scoring
• clip-criteria.md - What makes a good clip
• script-notes-template.md - Script notes format
• youtube-description-template.md - YouTube SEO template
• clipper/SKILL.md - Full clip extraction workflow
• x-post-extractor/SKILL.md - X post extraction workflow
• x-post-extractor/VOICE.md - Writing voice guidelines