AI Tutors Guide

A model plus a set of standing instructions

The model does the reasoning. The instructions decide what it is expert in, how it explains, how long its answers run, and what vocabulary and language it uses. Change the instructions and the same model behaves like a different teacher — which is why the brief matters more than the model choice for most subjects.

Why several narrow tutors beat one general one

• Each keeps its own vocabulary and tone, so you stop re-establishing context every session.
• Each can sit on a different model — a small fast one for drilling, a larger one for reasoning.
• You can hand a colleague one tutor without handing over everything else you have built.
• When one starts giving poor answers, you know exactly which brief to fix.

Step by step

1. Start a new tutor and name it for what it teaches and at what level, not just for the subject — you will have several.
2. Choose its model. A local model keeps the entire session on your machine; the optional Nexus AI module gives you a larger cloud model in exchange for Merits.
3. Write its instructions: the subject, the level of the learner, how much working to show, and what it should do when the learner is wrong.
4. Set the response length and the temperature. Keep the temperature low for anything with a right answer, higher when you want alternative explanations.
5. Test it with a question you already know the answer to, then fix the instructions rather than rephrasing the question.
6. Save it and reuse it. A tutor you rebuild each time is a prompt, not a tutor.

Writing a brief that works

Be concrete about behaviour rather than reaching for adjectives. "Ask what the student has already tried before giving any answer; show every algebraic step; never jump to the result" does far more than "be a helpful, patient tutor", because only the first one is checkable.

General academic support

• Break a concept into parts and check understanding at each one.
• Give a real-world example and an analogy for anything abstract.
• Ask questions back rather than handing over the answer.
• Adapt the depth of the explanation to the answers it gets.

Mathematics

• Algebra, calculus, geometry, trigonometry, statistics and probability.
• Insist on step-by-step working, with the reason for each step stated.
• Keep the temperature low — arithmetic does not benefit from creativity.
• Ask it to verify the answer a second way before presenting it.

Science

• Physics, chemistry, biology and the life sciences.
• Lab procedure, the scientific method, and reading data from a table or chart.
• Require units and significant figures in every numerical answer.
• Ask for the assumption behind a result, not just the result.

Anything else

The same pattern builds a languages tutor, a history essay coach, a code reviewer or an exam-technique drill. The subject expertise lives in the brief, not in the software, so there is no list of supported subjects to work around.

Load the book your class actually uses

A tutor otherwise answers from whatever a model absorbed during training. Load your textbooks, notebooks and syllabus documents into AI-Memory, and lesson plans, notes, slides, worksheets and question sets can then be generated from that material and mapped to your board's or university's scheme rather than to a generic curriculum.

Where this matters most

• Revision material that matches the chapter, the notation and the examples the class has already seen.
• Question sets pitched at the level the syllabus sets, not the level a model guesses.
• Marking handwritten answer sheets against your own key and rubric, on your own machine.
• Spotting the misconception a whole class shares, once a batch has been marked.

In practice

• Ask one question at a time. A stacked question gets a stacked, shallow answer.
• Paste in or photograph the actual problem instead of describing it from memory.
• Show your working and say where you got stuck — that is the material a tutor has to work with.
• Push back when an explanation does not land. Asking for a different analogy is a legitimate move, not a failure.
• Keep the good explanations by saving them into the project, so the session becomes your revision notes.

What to check before you rely on it

A model can be fluent and wrong at the same time, and it will not sound any less confident when it is. Check anything numerical, any date and any claim you would cite, and ask the tutor to show its reasoning so there is something concrete to check.