Yt To Blog

YT-to-Blog Content Engine

YouTube URL → blog post + Substack + tweets + vertical video clips. The whole content machine.

Pipeline Overview

YouTube URL
  ↓
① Transcript (summarize CLI)
  ↓
② Blog Draft (AI-written in your voice)
  ↓
③ Substack Publish (browser automation)
  ↓
④ X/Twitter Post (bird CLI)
  ↓
④b Facebook Group (optional reminder)
  ↓
⑤ Script Splitter (extract hook moments)
  ↓
⑥ HeyGen Videos (AI avatar vertical clips)
  ↓
⑦ Post-Processing (ffmpeg crop/scale)
  ↓
📁 Output Folder (blog.md, videos, tweet.txt, URLs)

One URL in → Five platforms out. Run the whole thing or any step individually.

First-Time Setup Wizard

Walk the user through this on first use. It takes ~10 minutes once, then never again.

Step 1: Check Dependencies

Run the setup script to check what's installed:

bash skills/yt-content-engine/setup.sh

Required CLIs:

If anything is missing, tell the user what to install and wait for confirmation.

Step 2: HeyGen API Key

1. Tell the user: "Go to https://app.heygen.com/settings — grab your API key from the API section."
2. If they don't have a HeyGen account: "Sign up at https://heygen.com — the free tier gives you a few credits to test with."
3. Save the key to config.json (see config schema below).
4. Test it:

curl -s -H "X-Api-Key: API_KEY_HERE" https://api.heygen.com/v2/avatars | python3 -c "import sys,json; d=json.load(sys.stdin); print('✅ API key works!' if 'data' in d else '❌ Invalid key')"

Step 3: HeyGen Avatar Setup

Tell the user:

"For vertical video clips, you need a HeyGen avatar. Here's what matters:

Record in PORTRAIT mode (hold your phone vertically). This is critical — if you record landscape, the avatar will be a small strip in the center of a 9:16 frame and we'll need to crop/scale it (which works but loses quality).

Go to https://app.heygen.com/avatars → Create Instant Avatar → follow their recording guide. Stand in good lighting, look at camera, speak naturally for 2+ minutes.

Once created, grab your Avatar ID from the avatar details page."

List their existing avatars to help them pick. Note: the avatars endpoint returns both custom and stock avatars — filter for the user's custom ones (they typically appear first and have personal names):

curl -s -H "X-Api-Key: API_KEY" https://api.heygen.com/v2/avatars | python3 -c "
import sys, json
data = json.load(sys.stdin)
for a in data.get('data', {}).get('avatars', []):
    print(f\"  {a['avatar_id']} — {a.get('avatar_name', 'unnamed')}\")
"

Step 4: HeyGen Voice Clone

Tell the user:

"Go to https://app.heygen.com/voice-clone → Clone your voice. Upload a clean audio sample (1-2 min of you speaking naturally). HeyGen will create a voice ID.

Once done, grab your Voice ID from the voice settings."

List their voices. User's cloned voices typically appear first; stock voices come after:

curl -s -H "X-Api-Key: API_KEY" https://api.heygen.com/v2/voices | python3 -c "
import sys, json
data = json.load(sys.stdin)
for v in data.get('data', {}).get('voices', []):
    print(f\"  {v['voice_id']} — {v.get('name', 'unnamed')} ({v.get('language', '?')})\")
"

⚠️ IMPORTANT: Use the FULL voice_id (e.g., 69da9c9bca78499b98fdac698d2a20cd), not a truncated version. The API will return "Voice validation failed" if you use a shortened ID.

Step 5: Substack Login

Substack has no API — posting requires browser automation.

1. Open the OpenClaw managed browser: use browser tool with profile="openclaw"
2. Navigate to https://substack.com/sign-in
3. Help the user log in with their credentials
4. Verify access by navigating to their publication dashboard
5. Save the publication URL to config.json

The browser session persists across restarts. One-time setup.

Step 6: Save Config

Create skills/yt-content-engine/config.json (relative to your workspace):

{
  "heygen": {
    "apiKey": "YOUR_API_KEY",
    "avatarId": "YOUR_AVATAR_ID",
    "voiceId": "YOUR_VOICE_ID"
  },
  "substack": {
    "publication": "yourblog.substack.com"
  },
  "twitter": {
    "handle": "@yourhandle"
  },
  "author": {
    "voice": "Description of your writing voice and style",
    "name": "Your Name"
  },
  "video": {
    "clipCount": 5,
    "maxClipSeconds": 60,
    "cropMode": "auto"
  }
}

Tip: If the user already has a voice guide from the yt-to-blog skill, read it from skills/yt-to-blog/references/voice-guide.md and use it for the author.voice field.

Step 7: Verify Everything

Run the setup script with the config in place:

bash skills/yt-content-engine/setup.sh

It will test each component and report status.

How to Invoke

Full Pipeline

"Turn this into a full content suite: https://youtu.be/XXXXX"
"Content engine this video: [URL]"
"Run the full pipeline on [URL]"

Individual Steps

"Just get me the transcript from [URL]"
"Write a blog post from [URL]" (steps 1-2)
"Post this to Substack" (step 3, after blog exists)
"Tweet about this blog post" (step 4)
"Generate video clips from this blog" (steps 5-7)
"Just split this into scripts" (step 5 only)

Pipeline Steps

Step ①: Transcript

Create the output directory for this run, then fetch the YouTube transcript:

mkdir -p /tmp/yt-content-engine/output-$(date +%Y-%m-%d)/scripts
mkdir -p /tmp/yt-content-engine/output-$(date +%Y-%m-%d)/videos

summarize "YOUTUBE_URL" --extract > /tmp/yt-content-engine/transcript.txt

The --extract flag prints the raw transcript without LLM summarization. Read the output. If it fails (no captions available), try with --youtube yt-dlp for auto-generated captions, or tell the user and suggest they provide a manual transcript.

Step ②: Blog Draft

Transform the transcript into a polished long-form blog post.

Load the author voice from config.json → author.voice. If a more detailed voice guide exists at skills/yt-to-blog/references/voice-guide.md, read and use that too.

Analysis phase — before writing, extract from the transcript:

• Core thesis — the single strongest argument or revelation
• Key data points — statistics, quotes, dates, names
• Narrative moments — anecdotes, examples, scenes
• Source links — URLs, studies, references mentioned
• Missing context — what does the reader need that the video assumed?

Writing structure:

1. Cold open (1-3 paragraphs): Scene-setting. Specific, sensory, emotional hook before data.
2. Thesis pivot (1 paragraph): Connect scene to the bigger story.
3. Data body (5-15 paragraphs): Alternate data and editorial. Each stat gets a punch line. Subheadings for major breaks only.
4. Callback (1-2 paragraphs): Return to opening scene/metaphor.
5. Closing (3-6 short paragraphs): Escalating fragments. Final hammer line.

Writing rules:

• Vary sentence length dramatically — long data sentences, then short punches
• Em dashes for asides, not parentheses
• Sentence fragments for emphasis
• No bullet lists in the body — narrative flow
• Inline source links, no footnotes
• No "in conclusion" or "to summarize"
• Credit video source naturally: "As [Name] put it..." with link
• Target: 1,500-3,000 words

Generate 3-5 headline options with distinct strategies (contrast/irony, revelation, moral framing, callback). Each with a subtitle. Let the user pick.

Save the final draft to the output folder as blog.md.

Step ③: Substack Publish

Post the blog to Substack via browser automation.

1. Read config.json → substack.publication
2. Open managed browser (profile="openclaw")
3. Navigate to https://PUBLICATION.substack.com/publish/post
4. Click the title field, type the title
5. Click the subtitle area, type the subtitle
6. Click the body area
7. Write markdown to a temp file, copy to clipboard (pbcopy < /tmp/post.md), paste into editor (Meta+v)
8. Substack auto-saves as draft

Known issues:

• Em dashes (—) may garble as ,Äî during clipboard paste → find/replace after paste
• Large posts: pause briefly between paste and verification
• Verify draft at https://PUBLICATION.substack.com/publish

Default: save as draft. Only publish if the user explicitly says "publish it" — always confirm first.

Save the Substack URL to output/substack-url.txt.

Step ④: X/Twitter Post

Compose and post using the bird CLI.

Compose the tweet/thread:

• If the blog has a single killer hook → single tweet with link
• If there are multiple strong points → thread (3-5 tweets)
• Include the Substack URL
• Match the author's voice but punchier — tweets are hooks, not summaries
• Use the handle from config.json → twitter.handle

Post with bird:

# Single tweet
bird tweet "Your tweet text here"

# Thread (post first tweet, then reply to it)
bird tweet "Tweet 1 text here"
# Note the returned tweet ID, then:
bird reply TWEET_ID "Tweet 2 text here"
# And chain:
bird reply TWEET_2_ID "Tweet 3 text here"

Always show the user the tweet text before posting and get confirmation.

Save tweet text to output/tweet.txt.

Step ④b: Facebook Group (Optional)

If config.json includes a facebook.group URL, remind the user to post to their Facebook Group.

Note: Facebook Group API posting is heavily restricted. Browser automation is unreliable due to Facebook's anti-bot measures. Best approach:

1. Draft a Facebook post version of the content (shorter, more casual than tweet)
2. Save to output/facebook-post.txt
3. Remind the user: "Don't forget to post to [Group Name] — here's your draft"
4. User posts manually

This keeps Facebook distribution in the workflow without fighting their API restrictions.

Step ⑤: Script Splitter

Extract 3-5 "hook moments" from the blog post and rewrite each as a spoken-word script for vertical video.

What to look for (scan the blog for these patterns):

1. Hook/Controversy — the most provocative claim, the thing that makes people stop scrolling
2. Data Bomb — a surprising statistic or fact that reframes understanding
3. Counterintuitive Take — something that contradicts conventional wisdom
4. Emotional Moment — a story, anecdote, or human element that creates connection
5. Call-to-Action Closer — a rallying cry, challenge, or "what you should do now"

Not every blog will have all five. Extract what's there. Minimum 3 clips.

Rewrite rules for spoken delivery:

• Hook first — open with the most attention-grabbing line. No preamble.
• Conversational — write for speaking, not reading. Contractions, natural rhythm.
• 30-60 seconds each — roughly 75-150 words per clip
• Self-contained — each clip must work on its own, no "as I mentioned earlier"
• End with punch — close on the strongest line, not a trailing thought
• No stage directions — just the words to speak, nothing else

Format each script:

CLIP 1: [descriptive title]
---
[Script text here, 75-150 words]

Use config.json → video.clipCount for the target number of clips (default: 5). Use config.json → video.maxClipSeconds for max duration (default: 60).

Save scripts to output/scripts/clip-1.txt, clip-2.txt, etc.

Step ⑥: HeyGen Video Generation

Submit each script to HeyGen API v2 to generate AI avatar videos.

Read config:

# Parse config.json
API_KEY=$(python3 -c "import json; c=json.load(open('config.json')); print(c['heygen']['apiKey'])")
AVATAR_ID=$(python3 -c "import json; c=json.load(open('config.json')); print(c['heygen']['avatarId'])")
VOICE_ID=$(python3 -c "import json; c=json.load(open('config.json')); print(c['heygen']['voiceId'])")

For each script, submit a video generation request:

curl -s -X POST "https://api.heygen.com/v2/video/generate" \
  -H "X-Api-Key: $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "video_inputs": [{
      "character": {
        "type": "avatar",
        "avatar_id": "'"$AVATAR_ID"'",
        "avatar_style": "normal"
      },
      "voice": {
        "type": "text",
        "input_text": "'"$(cat output/scripts/clip-1.txt)"'",
        "voice_id": "'"$VOICE_ID"'"
      }
    }],
    "dimension": {
      "width": 1080,
      "height": 1920
    }
  }'

Parse the response to get video_id:

import json
response = json.loads(response_text)
video_id = response["data"]["video_id"]

Submit ALL clips before polling. HeyGen renders in parallel — submit all scripts first, collect all video_ids, then poll them all. This cuts total render time from N×3min to ~3min.

Poll for completion (every 15 seconds, timeout after 10 minutes):

curl -s -H "X-Api-Key: $API_KEY" \
  "https://api.heygen.com/v1/video_status.get?video_id=$VIDEO_ID" \
  | python3 -c "import sys,json; d=json.load(sys.stdin)['data']; print(d['status'], d.get('video_url',''))"

Statuses: pending → processing → completed (with video_url) or failed (with error).

Download completed videos:

curl -L -o "output/videos/clip-1-raw.mp4" "$VIDEO_URL"

Credit note: ~1 credit per 1 minute of video. A typical 5-clip run uses ~3 credits. Warn the user about credit usage before submitting.

Step ⑦: Video Post-Processing

If the avatar was recorded in landscape (common), the 9:16 video will show a small avatar strip centered in a large frame with background fill. Fix this with ffmpeg.

Check config.json → video.cropMode:

• "auto" — detect and crop automatically
• "portrait" — skip cropping (avatar was recorded in portrait)
• "manual" — ask user for crop coordinates

Auto-crop pipeline:

# 1. Detect content bounds by scanning center column for non-background pixels
# Extract a single frame
ffmpeg -i input.mp4 -vframes 1 -y /tmp/frame.png

# 2. Use ffmpeg cropdetect to find content bounds
ffmpeg -i input.mp4 -vf "cropdetect=24:16:0" -frames:v 30 -f null - 2>&1 | grep cropdetect

# Parse the crop values from output: crop=W:H:X:Y

# 3. Crop content strip, scale up, center-crop to 1080x1920
ffmpeg -i input.mp4 \
  -vf "crop=DETECTED_W:DETECTED_H:DETECTED_X:DETECTED_Y,scale=1080:-1,crop=1080:1920:0:(ih-1920)/2" \
  -c:a copy \
  -y output.mp4

Alternative manual detection (preferred — cropdetect often fails when background is white/light):

HeyGen typically renders landscape avatars centered on a white/light background in the 9:16 frame. Scan the center column for non-white pixels to find the actual content strip:

# Extract a frame, then scan center column for content bounds
ffmpeg -y -ss 5 -i input.mp4 -frames:v 1 /tmp/frame.png 2>/dev/null

ffmpeg -y -i /tmp/frame.png -vf "crop=1:ih:iw/2:0,format=gray" -f rawvideo -pix_fmt gray - 2>/dev/null | \
  python3 -c "
import sys
data = sys.stdin.buffer.read()
first = last = None
for i, b in enumerate(data):
    if b < 240:  # Non-white pixel = actual content
        if first is None: first = i
        last = i
if first is not None:
    print(f'CONTENT_Y={first}')
    print(f'CONTENT_HEIGHT={last - first}')
    print(f'CENTER={( first + last) // 2}')
else:
    print('No content bounds detected — avatar may already fill the frame')
"

Then crop the content strip, scale proportionally to fill width, and center-crop to 9:16:

ffmpeg -y -i input.mp4 \
  -vf "crop=iw:CONTENT_HEIGHT:0:CONTENT_Y,scale=-1:1920,crop=1080:1920:(ow-1080)/2:0" \
  -c:v libx264 -crf 23 -preset fast -c:a aac \
  output.mp4

Proven crop values for common HeyGen landscape avatars (1080x1920 canvas):

• Content strip typically at y≈656, height≈607px
• Example: crop=1080:607:0:656,scale=3413:1920,crop=1080:1920:1166:0
• Always detect per-video — avatar placement can shift

Save processed videos to output/videos/clip-1.mp4, clip-2.mp4, etc.

If crop mode is portrait, just copy the raw files:

cp output/videos/clip-1-raw.mp4 output/videos/clip-1.mp4

Step ⑧: Output

Organize everything in a dated output folder:

output-YYYY-MM-DD/
├── blog.md              # Final blog post
├── tweet.txt            # Tweet text (posted or ready to post)
├── substack-url.txt     # URL of Substack draft/post
├── scripts/
│   ├── clip-1.txt       # Spoken word scripts
│   ├── clip-2.txt
│   └── ...
├── videos/
│   ├── clip-1.mp4       # Final processed vertical videos
│   ├── clip-2.mp4
│   └── ...
└── manifest.json        # Run metadata

manifest.json:

{
  "source": "https://youtu.be/XXXXX",
  "date": "2026-02-03",
  "blog": "blog.md",
  "substackUrl": "https://...",
  "tweetUrl": "https://...",
  "clips": ["clip-1.mp4", "clip-2.mp4", "..."],
  "heygenCreditsUsed": 3
}

Report the summary to the user:

• ✅ Blog post: X words
• ✅ Substack: [URL] (draft/published)
• ✅ Tweet: posted / ready to post
• ✅ X video clips generated and processed
• 💰 HeyGen credits used: ~X

Config Reference

Config file: skills/yt-content-engine/config.json (relative to workspace root)

Tips & Troubleshooting

• HeyGen rendering takes 2-3 min per clip. Set expectations — a 5-clip run takes 10-15 minutes of render time.
• Portrait avatars save time. No cropping needed. Worth re-recording if you use this regularly.
• Substack session expires? Re-run the browser login step (Step 5 of setup).
• bird CLI not posting? Run bird auth to re-authenticate.
• Bad crop detection? Switch cropMode to manual and eyeball the content bounds from a frame export.
• HeyGen quota errors? Check credits at https://app.heygen.com/settings — upgrade plan or reduce clip count.
• Transcript unavailable? Some videos don't have captions. Try summarize "URL" --extract --youtube yt-dlp for auto-generated captions, or ask the user for a manual transcript.