How I read papers in 2026

For years, the three-pass approach has been my go-to for reading technical papers.

But over the last 6-12 months, my process has changed completely. I've reshaped each pass into a personalized learning experience that matches how I actually think and absorb information.

This change happened naturally as I embedded AI agents and tools into all my workflows. Now, I read papers both faster and more effective than when I used to follow a one-size-fits-all approach.

Here's how my agentic three-pass approach works in 2026.

First pass

I don't read the abstract, skim through headings or look at figures to get the big picture anymore. Instead, I open NotebookLM, add the paper as a source, and ask it to generate a personalized summary:

From there I can just ask follow-up questions, save a few notes, and quickly decide if the paper is worth a second pass.

Second pass

This is where personalization starts to shine. I want to go a bit deeper into the paper: understand the methodology, related work, key arguments and main contributions. The goal is that I should be able to explain the paper's core contribution and methodology to a peer afterward.

Generate an infographic and slides deck

I'm a visual learner, so my first step is to generate an infographic. NotebookLM lets you tweak the generation instructions to adapt the output to the style you like the most but also to control the content that goes into that infographic.

Here is an example prompt I typically use:

Create an infographic with the key highlights of the paper and any relationships between them. In one corner, outline related work that led to this paper. Use a black background and a minimalist style.

If I have some preliminary understanding of the paper, I might also provide some more specific instructions in the prompt as well.

You can try to push this further, as much as you want, through instructions. Here is a more sophisticated example:

Turn this paper into a visual decision tree for practitioners. Start with the core problem at the top. Branch into the main methodological choices and trade-offs. Highlight failure modes in red and validated patterns in green. Add short, 1-sentence takeaways at each leaf node. On the right side, include a compact timeline showing how the idea evolved from prior work. Use a dark blue background, thin grid lines, and a clean mono font.

Pretty cool right? I must warn you that (for now) the more you try to include in the instructions, the more difficult it's for the model. I don't recommend overly complex instructions. Simpler prompts produce more reliable results.

Another feature I rely on during pass one is slide deck creation. I usually generate a deck covering the background, methodology, and conclusions of the paper. And you can customize this experience too!

Create a slides deck for AI engineers, with the background of the paper, main contributions, explanation of the methodology and the results. Use a minimalist style, with mono font and dark background. Include only the necessary text to explain the visuals.

Follow-up questions and selective reading

Once I have a solid high-level understanding of the paper, I start asking follow-up questions in the NotebookLM chat interface. I do this to gain a deeper understanding of concepts that I haven't fully grasped yet with the higher level materials.

You can also use another agent for this, like Claude: add the paper PDF to the context and start asking questions. The benefit I find in staying in NotebookLM is that I can easily save notes alongside other artefacts.

Occasionally, I might still want to cross-check the explanations against the original paper content. The source viewer in NotebookLM isn't great yet, so I usually open Zotero and read the relevant sections there.

Third pass

The third pass is essentially dev work, and it only happens when: 1) you're very curious about the paper and want to fully internalize it, or 2) the paper is directly relevant to your work and you want to implement everything, or parts of it. In practice, it's usually the latter for me due to time constraints.

At this point, I walk away from NotebookLM and switch to the terminal. I convert the paper to markdown, write my explanations and additions, and add this to a context folder in the repository where I'll be developing. This can be a either new repo or an existing one.

Something that has worked very well for me is to spawn a few instances of Claude Code within an existing codebase, and ask them to analyse if I could implement the paper's methodology to our solution, how much rework is needed, etc. and put together a plan.

Then, I review the plans to have good coverage over the range of possibilities. If I see a clear road ahead, I ask Claude Code instances to build a PoC and start testing it.

Conclusions

We've had ads and content recommendation for a while now. But highly personalized learning was still lagging far behind. In 2026, I think it's finally catching up.

Reading papers feels fundamentally different for me now. I can read and digest papers in the way I learn best, with every pass adapted to how my brain synthesizes information. And after that, I know I have an army of agents ready to help me put ideas to test in hours.

This combination creates a real flywheel: personalization makes learning faster, coding agents make experimentation cheaper, and together they remove the mental bottleneck that used to stop me from going deeper. You read a paper, you understand it on your terms, and you can actually go build or improve something with it the same day.

I see this area improving massively over the next few months. Image and video generation in tools like NotebookLM is still a bit rough, but I can already see how powerful those features will become.