微信扫一扫YouTube Content Intelligence Pipeline — UNATTENDED MODE
This is an automated invocation. There is no interactive user. All decisions must be made autonomously based on the thresholds below.
You are executing the /yt-intel pipeline. This is a 6-step process that extracts intelligence from a YouTube video, gates it for relevance to Agentic Delivery / Delivery Factory, and (if relevant) produces a branded report with Drive archival and Discord notification.
Input: A YouTube URL (e.g., https://www.youtube.com/watch?v=VIDEO_ID)
Step 1: EXTRACT — Video Metadata + Transcript
1a. Metadata extraction
Run via Bash:
yt-dlp --dump-json "<URL>" > /tmp/yt-intel/metadata.json
Parse from the JSON:
| Field | JSON key |
|-------|----------|
| Title | title |
| Channel | channel (or uploader) |
| Published date | upload_date (format: YYYYMMDD -> reformat to YYYY-MM-DD) |
| Duration | duration (seconds -> format as MM:SS or HH:MM:SS) |
| Views | view_count |
| Likes | like_count |
| Tags | tags (array -> comma-separated) |
| Chapters | chapters (array of {title, start_time, end_time}) |
| Description | description (first 500 chars) |
1b. Transcript extraction
# Download auto-generated subtitles
yt-dlp --write-auto-subs --sub-lang en --skip-download -o "/tmp/yt-intel/transcript" "<URL>"
This produces a .vtt file. Clean it:
# Parse VTT -> clean text (deduplicate, remove timestamps)
python3 -c "
import re, sys
with open('/tmp/yt-intel/transcript.en.vtt', 'r') as f:
content = f.read()
# Remove VTT header and metadata
lines = content.split('\n')
text_lines = []
seen = set()
for line in lines:
line = line.strip()
# Skip timestamps, empty lines, VTT headers, and style tags
if not line or '-->' in line or line.startswith('WEBVTT') or line.startswith('Kind:') or line.startswith('Language:') or line[0:1].isdigit() and len(line) < 10:
continue
# Remove HTML tags
clean = re.sub(r'<[^>]+>', '', line)
if clean and clean not in seen:
seen.add(clean)
text_lines.append(clean)
with open('/tmp/yt-intel/transcript_clean.txt', 'w') as f:
f.write(' '.join(text_lines))
print(f'Transcript: {len(text_lines)} unique lines, {len(\" \".join(text_lines))} chars')
"
Read the clean transcript for analysis. If transcript extraction fails (no captions available), continue with metadata-only analysis.
Step 2: ANALYZE — Extract Wisdom
Apply the Fabric extract_wisdom structured analysis to the transcript. Produce these sections:
Output Structure
# YouTube Content Intelligence Report
> **Generated:** YYYY-MM-DD | **Pipeline:** /yt-intel (extract_wisdom)
> **Source:** [Title](URL)
---
## Video Metadata
| Field | Value |
|-------|-------|
| **Title** | ... |
| **Channel** | ... |
| **Published** | ... |
| **Duration** | ... |
| **Views** | ... |
| **Likes** | ... |
| **Tags** | ... |
---
## Summary
[3-5 sentence overview of the video's core argument and conclusions]
---
## Ideas
[10-15 numbered key concepts/ideas, each 1-2 sentences with **bold lead**]
---
## Insights
[5-8 numbered deeper analyses — non-obvious takeaways, each 2-3 sentences with **bold lead**]
---
## Quotes
[8-12 notable verbatim quotes from the speaker, attributed]
---
## Facts
[10-15 numbered factual claims with specific numbers, studies, or data points]
---
## Habits & Practices
[5-10 numbered actionable habits or practices mentioned]
---
## References
[Table: Type | Name | Context — for all tools, people, companies, studies, products mentioned]
---
## Recommendations
[8-12 numbered recommendations from the speaker]
---
## One-Sentence Takeaway
> **[Single most important point in bold]**
Quality bar: Each section must have substantive content. If the video doesn't support a section (e.g., no specific quotes), write "None identified" rather than fabricating content. Facts must be verifiable claims from the video, not your own additions.
Step 3: RELEVANCE GATE
Score the video's relevance to Agentic Delivery / Delivery Factory on a weighted 1-10 scale.
Scoring Criteria
| Criterion | Weight | What counts as relevant | |-----------|--------|------------------------| | Dark factory / autonomous delivery | 3x | Software factories, lights-out development, autonomous agents building software, zero-human-code workflows | | AI coding tools & practices | 2x | Claude Code, Codex, Cursor, Copilot, agentic coding, vibe coding, spec-driven development, AI-assisted dev workflows | | DevOps / infrastructure automation | 2x | CI/CD, containers, monitoring, deployment automation, GitOps, infrastructure-as-code | | Enterprise software delivery | 1x | Engineering management, team structures, developer productivity, delivery metrics, technical leadership | | Business model / pricing / GTM | 1x | SaaS pricing, product-led growth, AI-native business models, developer tools market | | Irrelevant content | -5x | Gaming, lifestyle, politics, entertainment, cooking, sports, or other unrelated topics |
Calculation
- Rate each criterion 0-10
- Multiply by weight
- Sum weighted scores, divide by total weight (10)
- Clamp to 1-10 range
Decision Thresholds (UNATTENDED MODE)
| Score | Action |
|-------|--------|
| >= 6 | PROCEED — Continue to Steps 4-6 (DOCX, Drive, Discord) |
| 3 - 5 | PROCEED WITH REDUCED PRIORITY — Continue to Steps 4-6 but prefix Discord title with [MARGINAL]. The pre-filter already scored this above threshold, so complete the analysis. |
| < 3 | SKIP — Output the summary and score to stdout. Print: "Low relevance to Delivery Factory (score: X/10). Skipping Drive upload and Discord notification." Stop here. |
Relevance Section (added to report if proceeding)
After the One-Sentence Takeaway, add:
---
## Relevance to Agentic Delivery
**Score: X/10**
[Brief justification of the score — 2-3 sentences]
| Video Concept | Agentic Delivery Equivalent |
|--------------|---------------------------|
| [concept from video] | [how it maps to our product/positioning] |
| ... | ... |
### Sales Talking Points Extracted
1. [talking point derived from video insights]
2. ...
Step 4: GENERATE DOCX (only if relevance gate passes)
4a. Write Markdown report
Save the complete report (Steps 2 + 3 output) to:
/tmp/yt-intel/YYYY-MM-DD_<video-title-slug>_<channel-slug>_extract-wisdom.md
Slug rules: lowercase, hyphens for spaces, strip special chars, max 50 chars per slug.
4b. Convert to branded DOCX
Use the Pandoc MCP with the Agentic Delivery reference template:
mcp__pandoc__convert-contents
input_file: "/tmp/yt-intel/YYYY-MM-DD_<slug>.md"
output_format: "docx"
reference_doc: "/home/burhan/github/dev-delivery-factory-deployment/.claude/templates/AgenticDelivery_Document-Template_v2.docx"
output_file: "/tmp/yt-intel/YYYY-MM-DD_<video-title-slug>_<channel-slug>_extract-wisdom.docx"
Verify the DOCX was created:
ls -la /tmp/yt-intel/*.docx
Step 5: UPLOAD TO DRIVE (only if relevance gate passes)
5a. Copy files to accessible path
The Google Workspace MCP restricts file:// URLs to ALLOWED_FILE_DIRS ($HOME and /tmp). Files in /tmp/yt-intel/ are already accessible.
5b. Upload DOCX report
mcp__google-workspace__create_drive_file
file_name: "YYYY-MM-DD_<video-title-slug>_<channel-slug>_extract-wisdom.docx"
fileUrl: "file:///tmp/yt-intel/<filename>.docx"
folder_id: "1ISaW-3u8Kcrl_sfnGUR8pYfivlmazFzD"
user_google_email: "burhan@ocut.se"
Folder: 05-YouTube-Intelligence/Summaries/ (ID: 1ISaW-3u8Kcrl_sfnGUR8pYfivlmazFzD)
IMPORTANT: If create_drive_file uploads with wrong MIME type (e.g. text/plain), the DOCX will render as garbage on Drive. Verify the upload is application/vnd.openxmlformats-officedocument.wordprocessingml.document. If it fails, copy the DOCX to $HOME first and retry from there.
5c. Upload clean transcript
mcp__google-workspace__create_drive_file
file_name: "YYYY-MM-DD_<video-title-slug>_transcript.txt"
fileUrl: "file:///tmp/yt-intel/transcript_clean.txt"
folder_id: "1IZBTofol-NwDEPWRzpc_HOy8PVr2JsXy"
user_google_email: "burhan@ocut.se"
Folder: 05-YouTube-Intelligence/Transcripts/ (ID: 1IZBTofol-NwDEPWRzpc_HOy8PVr2JsXy)
5d. Get shareable link
mcp__google-workspace__get_drive_shareable_link
file_id: <uploaded DOCX file ID>
user_google_email: "burhan@ocut.se"
Save this link for the Discord message.
Step 6: POST TO DISCORD (only if relevance gate passes)
Channel
#ad-intel — Channel ID: 1475173738646470826
Message Format
**INTEL REPORT** — <Clickbait Title>
<2-3 sentence hook with the most provocative or surprising finding from the video. Make it compelling enough that someone scrolling Discord stops to read.>
**Key stats:**
- <stat 1 — most striking number/claim>
- <stat 2>
- <stat 3>
- <stat 4 (optional)>
**Relevance (X/10):** <One sentence on why this matters for Delivery Factory>
Full analysis: <Drive shareable link>
Source: <YouTube URL>
Constraints
- Max 2000 characters (Discord message limit)
- The clickbait title should be your own reframing, not just the video title
- The hook should make someone want to read the full report
- Stats should be specific numbers, not vague claims
- Do NOT use
>>>block quotes (renders poorly in Discord)
Send via Discord MCP
Use discord_discover to find the send message action, then discord_execute to post. The channel ID is 1475173738646470826.
Error Handling
| Error | Recovery |
|-------|---------|
| yt-dlp not found | Pipeline error — yt-dlp must be pre-installed |
| No subtitles available | Continue with metadata-only analysis |
| Pandoc conversion fails | Check pandoc is installed. Fall back to uploading the .md file |
| Drive upload fails | Check Google auth. Retry once. |
| Discord post fails | Log error and continue — report is still on Drive |
| Video is age-restricted/private | Log error and skip |
Working Directory
All intermediate files go to /tmp/yt-intel/. Create it if it doesn't exist:
mkdir -p /tmp/yt-intel
Clean up previous run files at the start of each invocation, but do not delete files from other videos (check filenames).