Using AI as a Learning Coach: Six Prompts That Beat Passive Tutoring

Why Most People Use AI Wrong for Learning

Ask the average ChatGPT or Claude user how they learn with AI, and they will describe a workflow that is genuinely useful but learning-shallow. Paste an article. Ask for a summary. Read the summary. Move on. That workflow saves time. It does not produce learning that sticks.

The cognitive science here is well established. The 2006 Roediger and Karpicke studies on the testing effect, replicated dozens of times since, show that retrieving information from memory produces stronger long-term retention than re-reading the same information, by margins that are not subtle. A single retrieval test produces roughly 60% recall after a week, compared to 40% for repeated reading. That gap widens with longer delays.

Summarisation is closer to re-reading than to retrieval. The model does the cognitive work, and you do the comprehension work passively. What you remember from a summary is mostly the parts that were already familiar. What you needed to learn most — the unfamiliar parts — slides through.

The fix is not to abandon AI for learning. It is to flip the prompt structure so the model becomes the coach instead of the encyclopedia. Six prompts, all grounded in established cognitive science, do most of the work.

Prompt 1: The Structured Plan

The first prompt is for the start of any new topic. The point is not to learn the topic — it is to map the territory before you walk into it.

> "I want to learn [topic]. I have [X hours] across [Y weeks]. Build a learning plan that progresses from beginner to advanced, with five-to-seven milestones. For each milestone, list (a) what I should be able to do at the end, (b) the two or three resources you would prioritise, and (c) the most common beginner mistake at that stage."

Why this works: the structure forces the model to articulate a curriculum, which exposes scope you would not have anticipated. Most beginners massively underestimate the breadth of a field and overestimate the depth of any single concept. Mapping the territory once, before diving in, is the lowest-cost intervention in the whole sequence.

What to do with the output: do not follow it religiously. The plan is a skeleton, not a script. The point is to have a coherent picture of what you are trying to learn before you start, so that everything you do can be located on the map.

Prompt 2: Active Recall Quiz

This is the core retrieval-practice prompt. Use it after every study session, every chapter, every video.

> "I just learned about [specific subtopic]. Quiz me with five questions of increasing difficulty. After I answer each one, tell me whether I am right, what I missed, and one follow-up question that probes the same concept from a different angle. Do not give me the answers in advance."

Why this works: the testing effect requires actual retrieval, not recognition. Multiple-choice tests are weaker than free-recall tests, because recognition can be done without retrieval. The prompt insists on free recall, then immediately offers feedback, which is the form of testing that produces the strongest retention gains in the literature.

The follow-up question is doing extra work. The 2008 Karpicke and Blunt studies on retrieval-induced learning show that the second-order question — "okay, you got that one, now apply it to a new context" — produces transfer learning that the first-order question cannot. Transfer is the test of whether you actually understood the concept versus pattern-matched the question.

Prompt 3: The Feynman Test

The Feynman technique, named for the physicist Richard Feynman, is the practice of explaining a concept in plain language without using its technical vocabulary. If you cannot, you do not understand it.

> "I am going to explain [concept] to you in plain English, as if you were a smart 12-year-old. After I finish, point out (a) where I used technical terms without defining them, (b) where my analogy broke down, and (c) one specific question you would still have if I were really 12."

Why this works: the prompt converts a vague self-test into a concrete one with three measurable failure modes. Most people, when self-testing, do not catch their own jargon, because the jargon is invisible to them — that is what jargon is. The model catches it because it is reading literally.

The "smart 12-year-old" framing matters. Easier framings (a 5-year-old) push you toward oversimplification. Harder framings (a peer) let you hide behind shared vocabulary. The middle target — articulate adolescent — is where the comprehension test is strictest.

Prompt 4: The Compare-and-Contrast Probe

This prompt is for any topic where you have learned more than one related concept and need to know whether you can tell them apart.

> "I have just learned about [Concept A] and [Concept B], which can be confused. Give me three scenarios where it would be easy to mistake one for the other, and ask me which one applies. Then explain the distinction in two sentences."

Why this works: the comparative-discrimination literature, going back to Eleanor Rosch's prototype theory work in the 1970s, shows that the boundary between related concepts is where understanding actually lives. You do not understand inheritance versus composition in software design until you can name when each applies. You do not understand sympathetic versus parasympathetic nervous activation until you can identify which one is happening in a specific situation.

The model is good at generating discrimination scenarios because it has been trained on enormous amounts of comparative material. Use it.

Prompt 5: The "Find My Blind Spots" Probe

Use this prompt after you feel confident on a topic. It is specifically engineered to surface what you do not know that you do not know.

> "I have been studying [topic] for [X time]. Based on what an expert in this field would consider essential, what are five things you would expect me to know that I have probably not encountered yet? For each, tell me why an expert would consider it essential and where I should learn it."

Why this works: this prompt operationalises the Dunning-Kruger curve in reverse. The 1999 Dunning and Kruger paper showed that beginners systematically overestimate their competence because they lack the meta-knowledge to recognise what they do not know. Asking the model to map your blind spots from the perspective of an expert is the cheapest available correction.

Treat the output skeptically. Models can hallucinate "essential" topics that are actually niche. Cross-reference at least one item against a textbook, a syllabus, or a recognised expert in the field before adding it to your study plan.

Prompt 6: The Spaced-Retrieval Refresher

The final prompt is for keeping what you have learned. It runs at intervals, not after every session.

> "Three weeks ago, I learned [topic]. Without giving me any answers, ask me to recall (a) the three most important concepts, (b) the most common mistake people make, and (c) how I would explain one of these concepts in plain language. After I respond, tell me what I missed."

Why this works: the spacing effect, documented across more than a hundred years of research starting with Hermann Ebbinghaus in 1885, shows that information retrieved at increasing intervals decays much more slowly than information retrieved repeatedly in a short window. The optimal spacing depends on how long you want to remember the material — for permanent retention, the intervals stretch from days to weeks to months.

The prompt does not need to be perfect. The crucial element is that the retrieval is happening at a delay, not the day after the original learning. Anki, RemNote, and other spaced-repetition apps automate this. If you are not using one, this prompt is the manual version.

What to Avoid

Two failure modes show up consistently when people try to use AI for learning.

Failure mode one: outsourcing the thinking. "Explain this concept to me" is not learning. It is consumption. The point of these prompts is to keep your brain doing the retrieval, the explaining, and the comparison. The model's job is to test, not to teach. If you are reading more than you are answering, the workflow has drifted.

Failure mode two: trusting hallucinations. Models confidently produce wrong information, especially in technical domains and especially about specific facts (dates, statistics, citations). For high-stakes learning, the rule is: any specific factual claim from the model needs a second source. The model is reliable as a coach. It is unreliable as a primary source for facts you have not verified.

Putting It Together

A single learning session, run with these prompts, looks roughly like this. Twenty-five minutes reading the original material. Five minutes generating questions for yourself with prompt 2. Ten minutes attempting the Feynman test in prompt 3. Five minutes on a discrimination probe with prompt 4. Total: 45 minutes.

A week later, prompt 6 runs in fifteen minutes. A month later, prompt 6 runs again, in ten. The total time investment for durable learning of a substantial subtopic is roughly two hours, spread across a month.

That is the version of AI-assisted learning that works. The summarisation workflow is faster, but it is not learning. It is information passing through.

---

*Sources: Roediger and Karpicke, "Test-Enhanced Learning," Psychological Science (2006); Karpicke and Blunt, "Retrieval Practice Produces More Learning Than Elaborative Studying with Concept Mapping," Science (2011); Hermann Ebbinghaus, "Memory: A Contribution to Experimental Psychology" (1885); Dunning and Kruger, "Unskilled and Unaware of It," Journal of Personality and Social Psychology (1999); Cepeda et al. meta-analysis on the spacing effect, Psychological Bulletin (2006). For the Feynman technique, see Richard Feynman, "Surely You're Joking, Mr. Feynman!" (1985).*

---

*Join 132,000+ professionals at How Do I Use AI for evidence-based AI tutorials and frameworks.*