How it works
Every lesson follows the same eight steps, in order, and skips any it does not need. Here is what you do in each one and why it is there, how answers are marked, and the research behind it.
Last updated
The sample lesson shows these steps with the free demo lesson's real questions and marking.
The steps of a lesson
1. Find your starting point
You answer three short questions about the topic. They are not marked, and "I don't know" is a useful answer.
Why: your plan starts from what you already know, not from the first page of a textbook.
2. Agree on a plan
You see the topic broken into small ideas, usually 4 to 7, in an order where each one builds on the ones before it. You start the lesson, or ask for changes first.
To make the plan, the tutor may first search the web about your topic. Each idea then links the pages it rests on, and the lesson's teaching and checks draw on what those pages say.
Why: small steps in a sensible order are easier to learn, and you can see what is coming.
3. Skip what you know
If your first answers suggest you already know an idea, you get one question about it before any teaching. A complete answer, given as "Fairly sure" or "Certain", saves the idea without teaching it. Otherwise it is taught like the others.
Why: your time goes to the ideas you need.
4. Learn one idea at a time
The tutor explains the idea with examples, often a worked example and then your turn, and asks you questions back. While an idea is being taught, you can ask your own questions.
For any idea, you can ask for an interactive picture, such as a chart with sliders. You write a prediction first, then explore. Along with the picture, the AI states the numbers it should show at two or three settings. The app runs the picture in a sealed frame, moves the sliders and compares. A picture with a wrong number goes back to the AI to be fixed. One that still fails is shown with a warning, never as checked.
Why: guessing first, then seeing how far off you were, tends to make the answer easier to remember.
5. Check it in your own words
Each idea ends with an open question. You write an answer in a few sentences and say how sure you are. There are no options to choose from.
If your answer misses a key point, the tutor explains the missing part again from another angle, then asks a new question. After three tries you can move on, and the final check asks about that idea again.
Why: writing the answer yourself shows what you can really recall. Choosing from a list can make you look surer than you are.
6. Map it from memory
With the lesson hidden, you write how the ideas connect, in your own words.
Why the lesson is hidden: recalling it is the practice. Looking back at the lesson would turn it into rereading, which does much less for memory.
7. A final check
You get fresh questions on every idea you were taught. An idea passes only if your answer covers all of its key points. The ideas you missed come back with new questions, for up to three rounds in all. Ideas still missed after that are saved anyway and come back for review sooner.
Why: a check on everything at the end shows what you can recall without the lesson in front of you.
8. Plan your practice
Each idea gets its first review date. The harder it was for you in the final check, the sooner it comes back: from the next day to three days or more. Then you write a short plan in the form "When…, I will…".
Why: a plan tied to a time and place makes it more likely that you come back.
How answers are marked
Every question has a few key points that a good answer shows, usually two to four. The AI reads your answer and says which of them it covered. The app works out the mark from that, the same way every time.
- In a lesson check: an answer that misses any key point is not a pass yet. The tutor explains the missing part again from another angle, then asks a new question. After three tries you can move on.
- In the final check: an idea passes only when every key point is covered. Any miss sends it to the next round, up to three rounds.
- In a review: all the key points is good, or easy if you were certain. Some missed is hard. Under half, or "I don't know", is again.
You always see which key points were counted, so you can judge the mark yourself. The tutor cannot change a mark in the conversation.
Saying how sure you were matters too, though the AI judges your answer without being told. A confident answer that turns out wrong is pointed out, because corrections like that tend to stick. An answer that covers every point and that you were certain of is spaced out further.
Reviews
After a lesson, your Today page shows the ideas that are due, with an estimate of how long they will take. Each review is a new question about an idea, typed from memory. It may ask you to explain the idea, apply it or compare it with another, more so as the memory gets stronger.
The mark sets the next review date. Passing an advanced idea also counts as practice for the simpler ideas inside it. A badly missed review brings forward the ideas that include it.
When the tutor cannot mark
If your allowance has run out, or the AI is paused, reviews carry on. You see the key points of a model answer and tick the ones your answer covered, and the app records the review from your ticks. Tick a point only if your answer really said it: marking yourself strictly keeps the dates right.
How review dates are chosen
The app uses FSRS-6, an openly published model of memory (the flashcard app Anki uses models from the same family). It estimates how likely you are to recall each idea, and schedules the next review for when that chance falls to your target. The target is 90% unless you change it in Settings.
After 200 spaced reviews, Settings can fit the model to your own history. The fit is kept only if it predicts your reviews better than the standard one.
The research behind each feature
Each feature follows a learning method with research behind it. The figures are from studies of the methods, not trials of this app.
| Feature | What it does | Evidence |
|---|---|---|
| Spaced retrieval practice | Reviews are typed from memory on dates set by FSRS-6 | Practice testing g = 0.61 (Adesope et al., 2017); spaced beats massed retrieval at g = 0.74 (Latimier et al., 2021) |
| Reviews that know the map | Passing an advanced idea credits the ideas inside it; a badly failed review pulls forward the ideas that include it | Adapted from Math Academy's FIRe |
| Mastery check | Fresh open questions on every idea; misses come back for up to three rounds | Mastery learning averages 0.52 SD, 0.64 at a 91–100% bar (Kulik et al., 1990); about half of Bloom's "2 sigma" (von Hippel) |
| Open answers, not options | Every check and review asks you to produce the answer, moving toward application and transfer as a memory strengthens | Options inflated scores by 21 points across 20 medical schools (BMJ Open) |
| Map from memory | You rebuild how the ideas connect with the lesson hidden | Retrieval beat concept mapping by d = 1.50 a week later (Karpicke & Blunt, 2011) |
| Course map and skipping ahead | Topics track planned ideas and mastery; ideas you know can be tested out | Accelerated students outperformed same-age peers by g = 0.70 (Steenbergen-Hu et al., 2016) |
| Mixed problem sets | Problems from different methods interleaved, weaker ones more often | Interleaved math practice d = 0.83 in a 54-classroom trial (IES) |
| Teaching moves | Concrete cases before the rule, worked examples then your turn, confusable ideas side by side, feedback that explains | Case comparison d = 0.50 (Alfieri et al., 2013); high-information feedback d = 0.99 vs 0.46 (Wisniewski et al., 2020) |
| Checked explorables | Predict, explore a verified interactive picture, then explain what you saw | Animations beat static pictures at d = 0.37 (Höffler & Leutner, 2007) |
| Calibration | You rate your confidence before grading; confident errors are flagged and re-tested | High-confidence errors are the most likely to be corrected (hypercorrection) |
| Difficulty band | Questions get harder above 90% session accuracy and more direct below 70% | Keeps practice challenging without turning it into guessing |
| A plan for practice | A "when…, I will…" plan after each lesson, a night's sleep before the first review | If-then plans d = 0.65 (Gollwitzer & Sheeran) |
| Scheduling fitted to you | After 200 spaced reviews, Settings can fit FSRS-6 to your own history, kept only if it predicts better | FSRS-6 log loss 0.3460 vs 0.3561 for FSRS-5 on ~727 million Anki reviews (srs-benchmark) |
The demo lesson
The free demo lesson, on compound interest, is scripted. It makes no AI calls and marks answers by looking for key words, so it is much cruder than the real tutor. Phrase your answers plainly there. It costs nothing and does not use your allowance. The sample lesson walks through it step by step.
Who decides what
The AI writes the words: explanations, questions, feedback, plans and pictures, and it says which key points each answer covered. The app makes the decisions: which step comes next, your mark, your review dates, and whether a picture is shown as checked. About the AI says more, including what the AI sees and what it keeps.
See a whole lesson first. The sample lesson shows the free demo lesson step by step, with no account needed. To ask for an invitation, email help@thoughtpantheon.com.