notebooklm-pipeline

name	notebooklm-pipeline
description	[ENTRY POINT] Full NotebookLM pipeline — research → generate video → download → watermark → trim → generate meta files → cover → prepare upload → publish. Single entry point for all end-to-end video tasks.

NotebookLM Pipeline — Master Entry Point

This is the single entry point for the complete Panda Video Automation pipeline. All work starts here. Every step is handled directly — no redirections.

Pipeline Overview

Research → Video → Download → Watermark → Trim
→ Meta (title/desc/tags) → Cover
→ Prepare Upload → Publish

---

Step 1: Research

1. Check NotebookLM installation & login
2. Ask user for research topic
3. Create notebook and start deep web research (nlm research start "<topic>" -m deep -n <notebook-id>)
4. Tell user ~5 min, come back later
5. When user returns: check status, import all sources (nlm research import <id>)
6. Generate summary via nlm describe notebook <id>

Step 2: Generate Video

Generate video overview in Simplified Chinese:

nlm video create <notebook-id> --format explainer --style auto_select --language zh_Hans -y -j

Report artifact info to user. Tell them ~5 min, come back to check.

Check status:

nlm studio status <notebook-id> -j

Proceed when status is completed.

Step 3: Download → Add Watermark → Trim Last 4s

3a. Download Video

nlm download video <notebook-id> -o input/video.mp4

3b. Add Watermark

Apply input/watermark.png to the bottom-right corner of the video (1280×720).

ffmpeg -i input/video.mp4 -i input/watermark.png \
  -filter_complex "overlay=1055:645" \
  -codec:a copy input/video-watermarked.mp4 -y
mv input/video-watermarked.mp4 input/video.mp4

3c. Trim Last 4 Seconds (remove NotebookLM bumper)

DURATION=$(ffprobe -v error -show_entries format=duration \
  -of default=noprint_wrappers=1:nokey=1 "input/video.mp4")
TRIM_TO=$(echo "$DURATION - 4" | bc)
ffmpeg -i "input/video.mp4" -t "$TRIM_TO" -c copy "input/video-trimmed.mp4"
mv "input/video-trimmed.mp4" "input/video.mp4"

ffmpeg required: brew install ffmpeg (macOS) or apt install ffmpeg (Linux)

Step 4: Generate Meta Files

4a. Generate Engaging Title

Generate a title directly from the artifact title — do NOT ask the user to confirm or choose. Decide it yourself and use it immediately.

• Fixed short title rule: Always use a short title ≤12 Chinese characters as the main title for all platforms. Never present alternatives or ask the user to pick.
• Simple & clear: Title must be easy to understand at a glance — plain everyday words, the core event/point named directly, no jargon, no clever ambiguity.
• Full version (≤80 chars) — optional for Bilibili; only used if platform needs a longer title, still decided directly, no user confirmation.

The decided short title goes into title.json and as the first line of description.md.

4b. Create title.json

echo '{"title": "<decided-short-title>"}' > input/title.json

4c. Create description.md

Format: first line = headline title, then ≤3 topic lines, each ≤50 characters.

<video-title>

视频看点：
1. <topic-1, ≤50 chars>
2. <topic-2, ≤50 chars>
3. <topic-3, ≤50 chars>

cat > input/description.md << 'EOF'
<video-title>

视频看点：
1. <topic-1>
2. <topic-2>
3. <topic-3>
EOF

Constraints:

• First line is the headline title
• All content must be Simplified Chinese — avoid English text in the description
• No emoji anywhere — no 📊, no symbols, no emoticons in the description text. Plain text only.
• At most 3 topics (1. 2. 3.)
• Each topic line ≤ 50 characters — concise punchline, not full sentences
• Use plain numbered list (1. 2. 3.), NOT circled numbers (① ② ③)
• Total content should be informative and engaging

⚠️ When passing to PVA upload commands, use actual line breaks (multi-line string), NOT \n escape sequences. Shell treats \n as literal text, not newlines. Example:

VIDEO_DESC="<video-title>

视频看点：
1. <topic-1>
2. <topic-2>
3. <topic-3>" npx pva bilibili upload

4d. Create tags.json

Pick ≤3 most relevant tags for the topic.

cat > input/tags.json << 'EOF'
{
  "tags": ["<tag1>", "<tag2>"]
}
EOF

Constraints: Max 3 tags. Most platforms (e.g. Bilibili) enforce this limit.

Step 5: Generate Cover Image

Use Python/Pillow to generate a minimalist academic cover:

python3 << 'PYEOF'
from PIL import Image, ImageDraw, ImageFont

W, H = 1920, 1080
img = Image.new('RGB', (W, H), color=(0, 0, 0))
draw = ImageDraw.Draw(img)

font_path = '/System/Library/Fonts/STHeiti Light.ttc'  # macOS
# Linux: '/usr/share/fonts/truetype/wqy/wqy-microhei.ttc'

title = '<title-line-1>'
subtitle = '<title-line-2>'
caption = '<caption>'

font_title = ImageFont.truetype(font_path, 130)
font_sub = ImageFont.truetype(font_path, 86)

BRIGHT_YELLOW = (255, 215, 0)
LIGHT_GRAY = (200, 200, 200)
STROKE_COLOR = (30, 20, 0)

# Compute positions and draw...
# (See full script in minimalist-academic-cover skill)

img.save('input/cover.png')
PYEOF

Cover saved to:

• input/cover.png

Step 6: Prepare for Upload

mkdir -p output/video output/spider
cp input/video.mp4 output/video/video.mp4
cp input/title.json output/spider/title.json
cp input/cover.png output/video/cover.png

Step 7: Publish

Upload to desired platform(s) via PVA CLI:

# One-time login per platform
npx pva weixin login
npx pva bilibili login
npx pva douyin login
npx pva kuaishou login

# Batch upload to all 4 platforms in parallel (single command)
V="$(pwd)/input/video.mp4" && C="$(pwd)/input/cover.png" \
  && T="$(cat input/title.json | jq -r .title)" \
  && G="$(cat input/tags.json | jq -r '.tags | join(",")')" \
  && D="$(cat input/description.md)" \
  && node_modules/.bin/pva bilibili upload --video "$V" --title "$T" --desc "$D" --tags "$G" --cover "$C" \
  & node_modules/.bin/pva douyin upload --video "$V" --title "$T" --desc "$D" --tags "$G" --cover "$C" \
  & node_modules/.bin/pva kuaishou upload --video "$V" --title "$T" --desc "$D" --tags "$G" --cover "$C" \
  & node_modules/.bin/pva weixin upload --video "$V" --title "$T" --desc "$D" --tags "$G" --cover  \
  & 

This launches 4 browser instances in parallel, one per platform. Each runs independently without interference.

---

Output Structure

After pipeline completes:

input/
├── video.mp4          (trimmed + watermark applied)
├── watermark.png      (watermark source image)
├── title.json         (title metadata)
├── description.md     (≤3 topics)
├── tags.json          (≤3 tags)
└── cover.png          (minimalist cover)

output/
├── video/
│   ├── video.mp4      (ready for upload)
│   └── cover.png      (ready for upload)
└── spider/
    └── title.json     (ready for upload)

---

Usage

User: "make a video about [topic]"

Agent — just do it, no unnecessary questions:

1. Step 1 → Research (create, deep research, import, summarize)
2. Step 2 → Generate video (background, tell user ~5 min)
3. Step 3 → Download video → Add watermark → Trim last 3s
4. Step 4 → Generate meta files (titles, description, tags)
5. Step 5 → Generate cover
6. Step 6 → Prepare for upload
7. Step 7 → Ask user which platform to publish to

User: "check progress"

Agent:

nlm studio status <notebook-id> -j

---

Status Reporting

When user asks for status/progress, invoke the pipeline-status skill and follow its checklist format. The pipeline itself does not own the report format.

Core Rules

• Non-Blocking: Launch operations and return immediately. Check status later with nlm research status / nlm studio status.
• No Unnecessary Confirmations: Proceed directly unless genuinely ambiguous.
• Watermark Always: Every downloaded video gets 熊猫视频自动化引擎 + https://panda.szhshp.org watermark in the bottom-right corner before trimming.
• Meta Files Always Required: Every pipeline run must generate title.json, description.md, tags.json, and cover.png in input/.
• Trim Always: Every downloaded video gets last 4 seconds trimmed.