The backlog problem
My to-read list grows faster than I can read. Some of it is noise; some is a 40-page research PDF I’ll never get to.
If it’s core to your job, read it properly – no shortcut works.
But if the material is only tangential, or you’re unsure you need it, or it’s just “nice to have,” deep reading is a bad ROI trade.
The solution: let an agent drive your summarizer
The usual move is to skim, or paste things into NotebookLM manually and generate summaries, infographics, audio overviews. That works, but it’s still hands-on – upload, prompt, tweak, repeat.
The better move: let Claude Code drive NotebookLM for you. The right format depends on you – some skim, some listen, some prefer videos. Good news: coding agents can personalize this at scale.
I retain visuals best, so I optimize for slide decks and infographics. If you learn by listening, swap in audio overviews. The workflow is the same.
What you need
- A coding agent – I use Claude Code, but Amp or any other works. (My case for why every person should be building with coding agents).
- A Google account with NotebookLM access.
- The
notebooklm-pyCLI (setup below).
⚠️ Heads up on sharing: NotebookLM notebooks aren’t publicly shareable – you can only share with specific Google accounts. This workflow is optimized for personal learning: you open the notebook in your own browser or phone later. If you need to share output with a team, export the deck or screenshot it (which is what I did for the examples below).
90-second setup
# 1. Install uv (if you don't have it)
curl -LsSf https://astral.sh/uv/install.sh | sh
# 2. Install notebooklm-py
uv tool install notebooklm-py
# or run ad-hoc: uv run --with notebooklm-py notebooklm --help
# 3. Authenticate with your Google account (opens a browser)
notebooklm auth
# 4. Smoke test — list your notebooks
notebooklm notebooks list
# 5. Point Claude Code at this folder and paste the prompt below.My actual prompt
Please create learning materials for me using notebooklm-py (https://github.com/teng-lin/notebooklm-py).
It is already authenticated; use "uv run" to execute any CLI commands.
Make sure to generate the following artifacts:
- slide-deck
- infographic
Don't download them—just give me a link to my notebook so I can review it from my browser or my phone later.
## Style Guide
Tone: practical, direct, no fluff. Write for experienced developers and technical leaders.
- Lead with "what you can do with this" not theory
- Short sentences, action-oriented language
- Assume technical baseline — don't over-explain fundamentals
- Structure around concrete use cases and outcomes
- Angle: clear value prop in first line, end with a question or takeaway
Visual style: clean whiteboard aesthetic — hand-drawn diagrams, muted earth tones (greens, browns),
simple icons, arrows showing flow between components. Think: architect sketching on a whiteboard,
not polished corporate slides.
Reference my style blog: https://kyrylai.com and info about me: https://kyrylai.com/about-me/.
Don't use these as sources for the notebook, but use them to help augment questions to NotebookLM.
## Input Materials & questions:
<paste URLs, papers, docs here>Examples with results
Let’s walk through 3 examples – each with the input material, the questions I asked, and the compressed presentation output.
Note: Since NotebookLM notebooks aren’t publicly shareable (see above), I’ve attached each deck as a single compressed image – all slides stitched together – so you can scan the output at a glance.
1. Composer 2 Technical Report:
https://cursor.com/resources/Composer2.pdf
This is an extremely useful piece. But since I don’t fine-tune models right now, I want to stay up to date without losing focus on what matters to me.
https://cursor.com/resources/Composer2.pdf
How could I re-use it in my practice?Output (Compressed presentations):
2. TRIBE v2 Human Brain Model
Outside my day-to-day, but fascinating — exactly the kind of thing the backlog eats.
https://ai.meta.com/blog/tribe-v2-brain-predictive-foundation-model/
- What does it mean for consumers?
- Could AI read our thoughts?Output (Compressed presentations):
3. Claude Mythos Preview
https://www-cdn.anthropic.com/08ab9158070959f88f296514c21b7facce6f52bc.pdf
Self-exploratory, 200+ pages of Claude Mythos testing.
https://www-cdn.anthropic.com/08ab9158070959f88f296514c21b7facce6f52bc.pdf
- Is it really that scary?
- What does it mean for the cybersecurity market?Output (Compressed presentations):
I get the gist of a paper in the time it takes to drink a coffee – and more importantly, I know which ones deserve a second cup.
Across these 3 examples: ~15 minutes total to generate and review all three decks, vs. ~4 hours of straight reading.
What’s been sitting in your backlog for three months? Try it on that one!