The fastest path to an answer is the slowest path to understanding.

The tension, stated plainly

Handing work to a model and handing your learning to a model are two different acts, even though from the outside they are typed the same way. When you ask Claude to draft a commercial proposal you already know how to draft, you are buying time. When you ask it to solve the derivatives exercise you do not understand yet, you are buying the feeling of having understood it. The first is good economics. The second is a debt that gets paid later, usually at the worst possible moment.

The trouble is that both feel equally good while they happen. Reading a clear explanation produces a sense of fluency that the brain reads as mastery. It is not. Mastery shows up when you recall the idea unaided, apply it to a case you have not seen, and notice when it does not apply. None of those three things happen while you read. They happen when you produce, and producing is uncomfortable.

Hence the working rule for this page: if the goal is to deliver, ask for the answer; if the goal is to learn, forbid it. This is not a moral position about technology. It is a design decision you make at the start of every session, and it changes how you configure the tool from top to bottom.

Making Claude ask instead of tell

A language model answers by default. It is trained to be helpful, and absent instructions to the contrary, helpful means solving. If you want it to interrogate rather than resolve, you have to say so explicitly and hold that line for the whole session, because the conversation drifts toward the answer the moment you show frustration.

These five instructions work, and they combine:

1. Forbid the solution in the opening instruction. Something as plain as Do not give me the answer under any circumstances, not even if I ask again or insist. Your job is to get me there. Anticipating the insistence is the load-bearing part: without that clause, it caves on the third request.
2. Commit before it speaks. Write your answer, your hypothesis or your reasoning first, and only then ask it to evaluate. Committing to a position before seeing the correction is what turns the correction into learning. If you read the answer before attempting it, the mistake never registers as yours.
3. Ask for the next question, not the solution. I am stuck here. Do not tell me what is missing. Ask me one single question that moves me closer. One. Give it three and the third contains the answer.
4. Reverse the roles. Explain the concept yourself, out loud or in writing, and ask it to play someone who does not get it and to question you wherever your explanation breaks. Teaching is the most honest exam there is, and the model is the only student available at eleven at night.
5. Ask for the error, not the fix. There is a flaw in my reasoning. Tell me which line it is on, without telling me what it is or how to fix it. Locating the error is enough information for almost everyone, and it keeps the cognitive work on your side.

What the five have in common is that they move the point of production. You still write the answer. The model only administers the difficulty.

A study system, not scattered questions

Asking loose questions in throwaway chats is the hack version of this: it works on Tuesday and by Thursday you cannot remember how you did it. Systems over hacks applies here too. A study system has three pieces and you assemble it once.

One Project per subject. The Project instructions are your teaching contract, and they apply to every conversation inside it without you repeating them: the level you are starting from, the language, the ban on direct answers, the format you want corrections in. Writing them well once saves you rewriting the same paragraph a hundred times and, more importantly, removes variation between sessions. You always study under the same rules.

The source material as context. Upload the chapter, the class notes, the official syllabus, past exam papers. The difference between asking a model about a topic and asking it about your topic, in your instructor's notation and within the real scope of your exam, is enormous. It stops giving you generic explanations and starts giving you explanations that match what you will actually be asked.

A Skill that drills you the same way every time. A Skill is a folder with a SKILL.md file that Claude loads when the task calls for it. It works by progressive disclosure: the description sits in context, the full file is read only when needed. For studying, that means you can write your own oral-exam protocol once (define the concept, apply it to a new case, say under what conditions it stops holding, find the error in this badly solved example) and run it identically every week. Repeating the protocol is what makes progress comparable.

For instructors: from syllabus to practice material

The expensive part of teaching is not explaining. It is manufacturing enough varied practice material and marking it. That is where a model genuinely changes the economics of the job, without touching the part only you can do.

From objective to exercise

Give it the learning objective, not the topic. "So they can distinguish correlation from causation in observational data" produces usable exercises. "Basic statistics" produces filler. Ask for ten variants of the same objective in different contexts and keep three.

Distractors with intent

On multiple-choice items, require every wrong option to correspond to a specific conceptual error, and have it tell you which. Then the exam result tells you what they misunderstood, not just how many got it wrong.

Rubrics before grades

Draft the rubric with the model, with levels described by observable behaviour rather than adjectives. Then have it apply your rubric to three real pieces of work and point out where the rubric is ambiguous. You are fixing the rubric, not the work.

Auditing your own material

Upload your slides and ask it to read them as a student at the actual level, not as a colleague: what does it assume you already know, which term appears undefined, at which jump does it lose you. It is the review nobody gives you because nobody has the time.

Solving like a mediocre student

Ask it to attempt your exercises while making the typical errors of the level. Ambiguous wording surfaces before the exam instead of during it.

None of those five decides what gets taught or why. That part is still yours, and it was the only part that mattered.

Assessment when the model can write the essay

The take-home essay stopped measuring what it used to measure. Lamenting that changes nothing, so it is better to go straight to what can be assessed instead. Everything below is workable with the resources you already have.

1. The process, not the deliverable. Ask for the trail: intermediate notes, discarded versions, the question that changed halfway through. Work with a history is hard to fabricate and easy to recognise.
2. Defence of the argument. Five minutes of conversation about their own text separates the person who thought it through from the person who commissioned it, with no detection tool involved. Ask why they discarded the obvious alternative.
3. Live reasoning. New problems, solved in front of you or within a bounded time, with the prompt "explain your next step before you take it".
4. Work with real sources. Hand them the source and assess what they did with it: what they cited, what contradicts their thesis, and how they resolved it.
5. Declared use. Require them to document what they asked the model and what they did with the response, as a graded component of the work. Making use declarable and assessable works better than banning it, because you cannot teach people to do well something you have forbidden.

Honesty rules to teach explicitly

Three are non-negotiable, and they are worth saying out loud, in class, more than once. First: fluency is not correctness. A well-written, confident, impeccably structured text can be wrong from beginning to end. The register of the prose carries no information about the truth of the content, and our instinct says otherwise.

Second: verifiable claims get verified. Dates, figures, quotations, attributions and bibliographic references leave the model and go into a search engine before they go into a piece of work. A perfectly formatted reference may not exist. Teaching the habit of going to the primary source is worth more than any warning about hallucinations.

Third: ask what would make it false. Telling the model "tell me what would have to be true for what you just asserted to be false, and how confident you are in each part" is one of the few cheap ways to separate what it knows from what it is completing. On its own, it is also a decent lesson in epistemology.

Where the model is genuinely strong for learning

After all that caution, fairness is due: there are things it does better than almost any other resource available, and they are not small.

It explains the same idea five different ways until one lands. A book has one explanation. A teacher, at best, two. Here you can ask for the geometric version, the version with small numbers, the version using an analogy from your own trade, the historical version of how it was discovered, and the version that starts from the case where it fails. One of the five is usually yours, and there is no way to know which in advance.

The patience does not run out. Asking the same thing for the fifth time carries a social cost in a classroom and carries none here. For a lot of people that cost is exactly what has stopped them asking for years, and removing it changes the outcome.

It works through the language barrier. A good share of serious technical material is in English. Being able to study it in your own language, ask for the original term alongside the translation, and practise the vocabulary of the field in both languages at once removes a toll that had nothing to do with the difficulty of the subject.

Accessibility and the size of the context

Vision matters more than it sounds in an educational setting. Diagrams, charts, a classmate's scanned notes, a PDF with mathematical notation, a photo of a whiteboard: all of it goes in and can be described, transcribed and explained. For someone with a vision impairment, or simply for someone handed a PDF that is an image with no selectable text, that capability is the difference between "inaccessible material" and "material".

The context window sets the other limit: how much material fits into a single study session. A million tokens is, in practice, a full textbook plus your notes plus the last five years of exam papers, all at once, with the model able to cross chapter 3 against question 12 of a paper from two years ago.

claude-haiku-4-5 · 200K

The fastest and cheapest. Flashcard review, marking short exercises, vocabulary drills, anything that repeats a lot and fits in a little. It is the only one of the four with a 200K context, not a million.

claude-sonnet-5 · 1M

The best speed-to-intelligence ratio. The daily study model and the one you will use for 80% of sessions, with the whole book loaded.

claude-opus-5 · 1M

The default choice for anything complex. Hard problems, designing a full course, deep review of your own material.

claude-fable-5 · 1M

The most capable widely released model, for the hardest reasoning. It requires 30-day data retention, so it is not available under a zero-retention policy. Worth knowing if you handle student data.

Extended thinking is tuned with an effort setting rather than a fixed token budget. Explaining a definition does not need it; taking apart a multi-step problem does. Raising it when the problem calls for it is cheaper than switching models out of habit.

What stays on the human side

Deciding what is worth learning. Holding up someone who has been stuck for three weeks. Noticing that the question being asked is not the question being had. Choosing what gets assessed and why. None of that is a context problem or a model problem, and none of it is going to move. What did move is the volume of mechanical work that surrounded it, and the sensible thing is to reclaim all of it for the list above. Clarity over noise: one subject, one system, and the production always on your side.