Issue Nº 006 · Revision

Revise the way the research actually says to

Re-reading your notes feels like work and barely moves the needle. Testing yourself, spaced out, is the method that does — and it used to mean hours of making flashcards by hand. AI collapses that to minutes. Here's the workflow, and the one place it will quietly lie to you.

With January exams close enough to feel, most students do the same thing: open the notes, read them front to back, highlight, and feel productive. It's the most popular revision method there is, and one of the weakest. The problem isn't effort — it's that recognising a page you've seen before is not the same as being able to produce the answer when the page isn't there. The methods that fix that are well established in the research. What's new is that the tedious part of using them is now free and takes minutes.

Two techniques do the heavy lifting, and they're not a productivity fad — they're among the most replicated findings in the science of learning. The first is retrieval practice: testing yourself instead of re-reading. The second is spacing: spreading that testing out over days rather than cramming it into one sitting. Everything below is just a way to get AI to do the boring half of both.

The evidenceTesting beats re-reading, and it isn't close

In a 2006 study by Roediger and Karpicke, students who read a passage once and then tested themselves on it remembered far more a week later than students who read the same passage four times. The re-readers felt more confident — and recalled less. A follow-up in Science in 2011 pitted retrieval practice against concept-mapping, a method students rate highly, and retrieval still won on a test taken a week on. The pattern is consistent: the harder, less pleasant study session is usually the one that sticks.

1 wk+
how far out the retrieval-practice advantage shows up — it's about long-term recall, not the same-day feeling
4×
times the re-readers in the 2006 study reviewed the material — and still remembered less than the self-testers

The catch has always been labour. Turning a module's worth of notes into good questions, then scheduling them so you see each one just before you'd forget it, is hours of admin nobody enjoys. That's the exact job a language model is now genuinely good at.

The workflowNotes in, a tested deck out

The whole thing is three steps, and you can run it on free tools.

The quality of the deck lives or dies on the prompt. A vague "make me flashcards" gives you bloated, two-paragraph cards that test nothing. This is the version worth saving:

The prompt

"You are helping me revise for an exam. From the notes below, write 20 active-recall flashcards. Rules: one fact per card; the front is a specific question, not a topic; the back is a short, exact answer. No cards that can be answered yes/no. Flag anything in my notes that looks incomplete or wrong instead of filling the gap yourself. Format as Question | Answer so I can paste it straight into Anki."

The "one fact per card" rule is the important one — atomic cards are what make spaced repetition work. The "flag, don't fill" rule is what keeps the model from inventing.

Where it slipsIt will confidently test you on things that aren't true

Here is the failure that actually costs you marks: a model asked to generate questions from your notes will sometimes answer from its own training instead — and when your notes are thin, it fills the gap with something plausible and wrong. You then drill that wrong answer, with spaced repetition, until it's burned in. The tool designed to make things stick will happily make a fabrication stick.

Retrieval practice is a method for remembering. It has no opinion on whether what you're remembering is correct.

Three habits defuse it. Keep the model grounded in your sources — this is why NotebookLM, or a prompt that says "only from the notes below," beats a bare chatbot. Spot-check the back of every card against your notes before you commit the deck; it takes five minutes and it's the five minutes that matters. And when the model flags a gap, fill it from the lecture or the textbook yourself — that gap is a real hole in your revision the AI just found for you, which is arguably the most useful thing it did.

None of this replaces understanding the material — it's for locking in what you already half-know so it's there under pressure. Done right, you spend your revision answering questions instead of re-reading pages, which is both the harder thing and the one that works.

✦ End of issue ✦

Get the next one in your inbox

One AI tool, one prompt, three quick wins, every week. Free, and under 3 minutes to read.

Subscribe Free →