notebooklm-overlay

name

notebooklm-overlay

description

Apply the locked-in Unite-Group editorial overlay system to any NotebookLM-generated video (Video Overview output, daily briefings, audio overviews rendered as video). Use AFTER video-use produces a base render; this skill applies a post-process editorial overlay pass that gives every NotebookLM video a distinctive house identity. Triggers on "NotebookLM video", "daily briefing video", "audio overview", "video overview", or any video with NotebookLM as the source aesthetic. Brand-agnostic: takes a beats JSON config (window/title/body per beat), renders editorial-serif cards (Charter italic title, Charter regular body, hairline rule, warm-cream card on dark hairline at the top edge) and composites them onto the bottom-third with 300ms cubic fade in/out. Composes with the video-use skill — video-use cuts and renders the base, this skill polishes.

notebooklm-overlay

The locked-in Unite-Group house identity for every NotebookLM-source video. Use this skill after video-use produces a base render to add the editorial overlay polish that distinguishes a Unite-Group NotebookLM video from a generic NLM output.

When to use

Fire this skill in any of these cases:

• Editing a NotebookLM Video Overview output
• Editing a daily briefing video (the Pi-CEO daily briefings are NLM-generated)
• Editing any video whose source aesthetic is NotebookLM's hand-drawn-illustration style
• The user mentions "NotebookLM video", "daily briefing video", "audio overview", "video overview"

Composition with video-use

video-use → preview_base.mp4   (rendered with --no-subtitles)
   ↓
notebooklm-overlay → preview_final.mp4   (editorial cards composited)

Important: Render the video-use base with --no-subtitles. The bold-CAPS subtitles in video-use's default bold-overlay style fight the editorial overlay tonally — pick one or the other. The editorial overlay wins for NLM videos because the cards carry the copy.

master.srt from video-use is still useful — keep it as a sidecar for YouTube / LinkedIn upload.

Usage

# Author beats.json aligned to the EDL output timeline
cat > beats.json <<'EOF'
{
  "beats": [
    {"window": [0.5, 6.5],   "title": "May 2026", "body": ["Pi-CEO running 11 portfolio businesses, zero-touch."]},
    {"window": [8.0, 19.5],  "title": "Swarm Activated", "body": ["Guardian  ·  Builder  ·  Scribe  ·  Click"]},
    {"window": [21.0, 47.5], "title": "May 2026 — What Shipped", "body": [
      "98 features across 12 sprints.   85 / 100 health score.",
      "Marathon-4:  a flawless six-hour autonomous run."
    ]},
    {"window": [49.0, 61.0], "title": "Powered by Claude", "body": ["1-hour prompt cache  ·  −80% tokens  ·  pre-compact hook"]}
  ]
}
EOF

# Render
~/Developer/video-use/.venv/bin/python ~/.claude/skills/notebooklm-overlay/overlay.py \
  --input preview_base.mp4 \
  --beats beats.json \
  --output preview_final.mp4

The script uses video-use's venv because it ships PIL pre-installed there. If you don't have video-use installed yet, install it first (uv sync inside ~/Developer/video-use/).

Locked house style (do not vary without explicit user approval)

token	value
card BG	rgba(250, 249, 246, 235) — warm cream, ~92% opaque
top rule	rgba(40, 40, 40, 220), 2px hairline
title	Charter italic 42pt, near-black
body	Charter regular 24pt, soft dark gray
position	bottom-third (y = video_height − card_height)
card height	220px (≈20% of 1080p frame)
pad-x	96px (5% of 1920)
fade	300ms cubic in / 300ms cubic out at each window edge

The visual identity is locked. The CONTENT (beats / titles / body lines / windows) is per-video.

beats.json schema

{
  "beats": [
    {
      "window": [start_s, end_s],   // numbers, seconds in the OUTPUT timeline
      "title": "string",             // serif italic, single line
      "body": ["string", "string"]   // 1–3 lines, serif regular
    }
  ]
}

Notes:

• window is in OUTPUT seconds (after video-use's cut), not source-clip time.
• Card auto-fits — keep body lines under ~70 chars each at 24pt to stay inside the 1920×220 card.
• For 4–5 beat reels, expect window durations of 6s (HOOK/UPGRADES) through 26s (METRICS).

Brand-agnostic by design

The same renderer handles any portfolio brand or client by swapping the beats.json content. Don't introduce per-brand colors / fonts unless the user explicitly asks for a brand-specific variant — the locked house style is the point.

If a future variant is needed (e.g., a particular client demands their own typography), fork overlay.py into a new skill (e.g., notebooklm-overlay-acme), don't mutate this one. Lock-in here means stable.

Anti-patterns

• Don't apply subtitles in the base render — they fight the editorial overlay.
• Don't position the card outside the bottom-third — NotebookLM puts its content in the upper 2/3.
• Don't use card BG with <80% opacity — body text becomes unreadable over busy NLM frames.
• Don't add per-brand colors / logos without an explicit user request — that breaks the "brand-agnostic" lock.
• Don't re-encode the audio when applying overlays — -c:a copy preserves the loudnorm pass from video-use.