The method: how The Learning Harness teaches

The rules the harness runs, in the order a learner meets them, and the evidence each one leans on.

By Ojasvi YadavPublished 6 September 2026

One learner, one record

The harness keeps everything it knows about you in one database inside your study folder. The record holds the concepts you have, the answers you gave when it placed you, the topics on your path and their order, every quiz and explain-back with its result, the questions you asked while learning, and the cards it wrote. Every lesson, quiz and grade reads from that record and writes back to it. Nothing depends on a chat transcript, so a new session opens exactly where the last one ended, and a lesson written next month is pitched at what you hold next month.

This is the property a chat window lacks and the reason the harness is an agent rather than a prompt. The four rules that follow only work when the AI can remember what happened yesterday. The reasoning is laid out in how to learn with AI, properly.

Step 1. It asks what you know

Before the first lesson of a topic, the harness lists the concepts that lesson will lean on and asks about them, one at a time. Three broad questions come first, because three is enough to place you without fatigue. Only if real gaps remain does it ask a second round of up to six targeted questions, each aimed at the exact uncertainty the first three exposed. The questions ask what you hold and in which vocabulary; they never hint at an answer and never turn into a test.

Every answer is stored against the concept it concerns. Later lessons, the lesson reviewer, and the quiz all inherit the same prior, so you are asked once.

Step 2. One ordered path, smallest step next

Every topic sits on one ordered path. Two rules decide the order. The hard rule: nothing is taught before its prerequisites, because a lesson with a missing prerequisite cannot be taught. The soft rule: among the lessons that are teachable, the next one is the smallest step from what you already hold.

What you want to learn falls into one of three cases, and the path handles each. Some of it is already inside what you hold, and the lesson connects concepts you have. Some of it is one or two concepts away, and the lesson teaches those first. Some of it is far off, and the harness builds a short course in order over several days, each lesson the smallest step from the last.

Step 3. A lesson written for you, reviewed before you read it

The harness writes each lesson as a self-contained page for one reader. It fetches sources for the lesson and cites them. It uses the prior from Step 1, carries one worked example through every part, draws a diagram wherever a mechanism has parts, and uses the field's own terminology so that later reading in the field feels familiar. It renders the page and looks at it before it hands it over, because a diagram that overlaps its labels teaches nothing.

A second AI pass then reviews the lesson against the sources, the prior, and the workspace's prose rules, and fixes what it finds. Only the reviewed page reaches you. A randomized trial at Harvard found that an AI tutor built around teaching practices produced more than double the learning gains of an active-learning class, in less time (Kestin and colleagues, 2025); the lesson rules above are the harness's version of those practices.

Step 4. A quiz the same day

The quiz starts after you say you have finished reading, never in the same message as the lesson. Each question asks for a mechanism, a prediction, or a boundary condition, one at a time, and never for yes or no. The harness grades each answer against the lesson it wrote and records the result per concept. Practice testing is one of the two techniques rated high utility in the largest review of learning techniques (Dunlosky and colleagues, 2013), and a same-day test raised one-week retention above rereading in Roediger and Karpicke (2006).

A passed quiz moves the concept to "currently learning" and schedules the explain-back for the next day. A reminder arrives the next morning if you have reminders switched on.

Step 5. An explain-back the next day, after sleep

The next day, with the lesson closed, you explain each concept in your own words and the harness grades the explanation against the lesson. It marks the mechanism, not the vocabulary. The delay is deliberate: after five minutes, restudying beats testing, and after two days the order reverses, so a same-day test cannot tell short-term memory from learning. A night of sleep consolidates declarative memory (Diekelmann and Born, 2010), and the next-day test measures what survived it.

A concept that fails is taught again with a fresh lesson and tested again the next day. It is never marked understood on the strength of the same-day quiz alone. The full protocol, with a prompt for people who use a general chat model, is in the Feynman technique with AI.

Step 6. Cards into Anki, ordered across concepts

Once a concept passes both tests, the harness writes two to four cards for it into Anki: the definition, the trap, one mechanism or number, plus one card for every question you asked while learning it. Each card asks one question with one gradable answer, in the field's official terms, and stands alone. Cards for a question you asked keep your own wording on the front, because the question you actually had is the best prompt for the answer that resolved it.

The harness orders Anki's queue of unseen cards so that each day spans several concepts instead of burying you in one, and it paces new material so that daily reviews stay near 50 cards. From then on it reads the review history, the intervals and lapses of each card, as the measure of retention, and it uses that history when it judges later quizzes. Distributed practice is the second high-utility technique in Dunlosky's review, and spaced study beat massed study at every interval in Cepeda and colleagues (2006). The card rules are in spaced repetition with AI.

What the harness refuses to do

It will not give you the answer during an exercise; it gives hints in steps and grades the attempt. It will not quiz you in the same message as the lesson. It will not write a card for a concept that has not passed both tests. It will not teach a topic before its prerequisites, and it will not mark a concept understood because you say you know it: a claim of prior knowledge earns a short verification quiz, and the explain-back still follows. Each refusal removes a way to feel finished without being finished. The field experiment by Bastani and colleagues (2025) measured what the first refusal is worth: a hint-only AI removed a 17% exam loss caused by an AI that handed over answers.

What it runs on

The harness runs in a terminal on macOS, inside a study folder you choose; Linux works with some gaps. One install command sets up what is missing: Node, Python, git, Anki desktop, the AnkiConnect add-on, and a hidden browser the harness uses to look at the lesson pages it writes. Your lessons and sources are ordinary files in the folder, kept in git, and your record is one database file beside them. Cards live in Anki, so they remain yours if you stop using the harness. It is in a closed beta; the FAQ has the details.

Sources

1. Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning. Scientific Reports, 15, 17458. https://doi.org/10.1038/s41598-025-97652-6
2. Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). Improving Students' Learning With Effective Learning Techniques. Psychological Science in the Public Interest, 14(1), 4–58. https://doi.org/10.1177/1529100612453266
3. Roediger, H. L., & Karpicke, J. D. (2006). Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention. Psychological Science, 17(3), 249–255. https://doi.org/10.1111/j.1467-9280.2006.01693.x
4. Diekelmann, S., & Born, J. (2010). The memory function of sleep. Nature Reviews Neuroscience, 11, 114–126. https://doi.org/10.1038/nrn2762
5. Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T., & Rohrer, D. (2006). Distributed practice in verbal recall tasks: A review and quantitative synthesis. Psychological Bulletin, 132(3), 354–380. https://doi.org/10.1037/0033-2909.132.3.354
6. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning. Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.2422633122