How to handle forty readings

If your degree is a pile of PDFs, the bottleneck is not reading speed. It is that you read forty things, kept none of them findable, and cannot answer the one question that matters across a whole module: where do these authors disagree. The setup that fixes it is free and takes about fifteen minutes once, in week one, and then runs itself for three years.

The five steps, in order#

1. A reference manager catches it. Install Zotero and its browser button. One click on any paper, chapter or news article saves the reference and the PDF together, correctly, permanently. This is the step everyone skips and later regrets, and the regret arrives in third year at the worst possible moment.

2. One folder per module, not per essay. A collection for the module means everything you read this term lives in one place whether you used it or not. Per-essay folders throw away the reading that did not make the cut, which is the reading you will want next year.

3. A notebook tool reads the pile. Upload that module's PDFs into a single notebook and ask questions across all of them at once. The important property is that it answers from your uploads and cites the line, so you can check it in a second. Still check it.

4. Ask the cross-reading questions. Where do these two authors disagree? Which of these supports my argument and which undermines it? What is nobody in this pile talking about? None of those can be answered by reading each item in turn.

5. Turn the pile into revision. The same notebook makes a quiz from your own readings, and an audio overview you can listen to walking to campus.

Why the tool that cites the line is a different kind of tool#

This distinction makes the workflow safe, so it is worth being precise about. A chatbot answering from its training fills gaps with something plausible. A notebook grounded in your uploads points you back to the page.

That matters more than it sounds. An audit across 2.5 million biomedical papers found 4,046 fabricated references in 2,810 papers, with the rate of papers carrying at least one rising from 1 in 2,828 in 2023 to 1 in 277 by early 2026. Review articles ran 57 per cent higher than other paper types. Those are published, peer-reviewed papers. A tool that can only quote documents you supplied cannot invent a source, which removes the single most damaging failure mode from your reading workflow.

It does not remove the others. It can still misattribute which of your uploads said a thing, and it can still summarise a limitation out of existence. That is what the third move of the source rule guards against: understand, rather than extract.

Step five is the one with the strongest evidence behind it#

Making a quiz out of your own readings is the single best-evidenced study technique in psychology, and most students default to the opposite.

Roediger and Karpicke showed that the winner reverses with delay. At five minutes, restudying beat testing, 81 per cent against 75. At one week, testing beat restudying, 56 per cent against 42. In their second experiment repeated study led at five minutes, 83 against 71, and trailed badly at one week, 40 against 61. Rereading feels better and is worse.

Karpicke and Blunt then put retrieval practice against concept mapping: 0.67 against 0.45, roughly a 50 per cent advantage in long-term retention, and 101 of 120 students did better after retrieval practice than after elaborative study. A published comment disputes the fidelity of the concept-mapping condition, which anyone citing it should say.

So the ordering is not arbitrary. Steps one and two make the pile findable, step three and four make it answerable, and step five is the one that puts any of it in your head.

The failure this design is built around#

Reading a summary feels like learning, and the feeling is the problem rather than the summary. Fluent material feels learned, which is the illusion of competence, and a pile of AI summaries is the most fluent version of a module you will ever encounter.

The workflow above is arranged so that the summarising is never the last step. Something has to test you afterwards, because the gap between feeling prepared and being prepared is invisible from the inside and only shows up when somebody asks.

What this has not been shown to do#

Nobody has tested this five-step workflow against any other, or against no system at all. The retrieval practice evidence is strong and it is about prose recall in a laboratory rather than about a notebook tool making quizzes from your uploads. The claim that grounded tools cannot fabricate is a claim about architecture, not a measured error rate, and no published study has compared hallucination rates between a grounded notebook and a general chatbot on student reading.

The tool names will also date faster than the sequence. Catch, organise, ground, cross-read, test is the part worth keeping.

Key sources

• Roediger, H. L. and Karpicke, J. D. (2006). Test-Enhanced Learning. Psychological Science, 17(3).
• Karpicke, J. D. and Blunt, J. R. (2011). Retrieval Practice Produces More Learning than Elaborative Study. Science, 331(6018).
• Topaz, C. M. et al. (2026). Fabricated citations: an audit across 2.5 million biomedical papers.

Related SuperSkills research#

On checking what you find, the source rule. For students, how to use AI at university, when your department bans it and how to be honest about it. On why testing beats rereading, retrieval practice and desirable difficulty. On the trap, should I let AI summarise everything I read.

About this research#

The five-step reading workflow is used by Rahim Hirji in teaching and appears in the Mastering AI deck, most recently in September 2026. No claim of first use is made for it, and cross-reading is described here as a practice rather than claimed as a coinage. Zotero and Gemini Notebook are named products and belong to their makers. Findings are attributed to the studies that produced them and kept separate from the interpretation.