2 Personalised tutor
This chapter will teach you how to use AI as a personalised tutor to explain concepts and functions you’re struggling to understand, or would like more information on. AI can act as a study partner to explain concepts in multiple ways, but it is not a replacement for your own practice or your lecturer’s guidance. Think of it as a supplement, not a substitute.
It will be helpful to work through these activities with a specific week/chapter/lecture of your course in mind.
2.1 Custom instructions
The key to maximising usefulness is to set up custom instructions (sometimes called “system prompts”). These apply across all chats so the AI knows who you are and how you want answers framed. This is about the level, format, and tone of the answers, it is not about “learning styles”: people do have preferences about how information is presented, but matching teaching to those preferences doesn’t improve learning, so don’t tell it you’re a visual learner.
These are mine. They don’t always get followed but the answers are much better than without them. Notice that most of it isn’t about who I am, it’s about how I want it to behave, and in particular that I’ve told it not to flatter me. AI is sycophantic by design: it agrees with you and tells you your work is brilliant about 50% more often than a human would, and people prefer it that way. You have to actively switch that off, and even then it only half works.
Coding. Code primarily in R using tidyverse solutions. Use Qmd files.
Tone and style. Formal and concise. Depth where the substance requires it, never padding. Never be sycophantic. Do not open with praise or affirmations. Act as a critical friend: disagree, question assumptions, and push back where warranted. Do not play devil’s advocate without good reason. Do not use mannered expressions.
Evidence. For academic writing and critique, provide peer-reviewed evidence preferably from the last five years. Fact-check claims against the literature and present alternative viewpoints where they exist.
Critique. When asked to critique, do not simply validate. Check claims against peer-reviewed evidence and present competing perspectives even when you agree with the position being critiqued.
The evidence and critique sections are there because I use it for research. You won’t need those yet, but the structure works at any level: a short section on who you are and what you’re using, then sections on how it should behave. Here are two versions written for students. The first is for someone in their second year, the second for someone starting out.
About me. I am a second year psychology student at the University of Glasgow on the course Applied Data Skills. I learned R in first year and I have reasonably good general computer literacy. I am using R and RStudio on a Mac.
Coding. Code in R using tidyverse solutions. Use Qmd files. Do not use base R when there is a tidyverse way of doing it. Always give a concrete, runnable example.
Tone and style. Concise. Never be sycophantic. Do not open with praise or affirmations. Do not tell me my question is a good one. Tell me when I have got something wrong.
Tutoring. Act as a tutor, not an answer machine. When I ask about a concept or function, explain it at the level of someone who has done one year of R. When I bring you an error or a problem, ask me what I think is wrong before you tell me.
About me. I am a first year psychology student at the University of Glasgow and I am learning R for the first time. I have never learned a programming language before and I am not very confident with computers. I am using R and RStudio on a Windows laptop.
Coding. Code in R using tidyverse solutions. Use Qmd files. Always give a concrete, runnable example and explain what each line does.
Tone and style. Plain language, short answers. Never be sycophantic. Do not open with praise or affirmations. Do not tell me my question is a good one. Tell me when I have got something wrong.
Tutoring. Act as a tutor, not an answer machine. Explain concepts as if I have never programmed before, but do not skip the technical terms; define them when you first use them. When I bring you an error or a problem, ask me what I think is wrong before you tell me. Do not give me the answer straight away.
Whatever else you write, the “never be sycophantic, do not open with praise” line is worth stealing, and so is “ask me what I think is wrong before you tell me”.
Write your custom instructions and enter them into Copilot. Exactly what information you provide is up to you but make sure that you explain your level of knowledge, skill, confidence and previous experience, and tell it how to behave (at minimum: don’t flatter me, and don’t give me the answer straight away). If you’re using it for coding, you also want to give it some technical information about the software you’re using (e.g., R and RStudio) and your operating system (Windows).
- To add your custom instructions in Copilot, click the three dots (…) in the top right corner, then Chat settings, then Personalisation, then Edit instructions under Custom instructions. Paste in your instructions and click Save instructions. Make sure the Custom instructions toggle is switched on. Ignore the suggestion buttons underneath the box (“Prioritise my manager” and so on), they’re aimed at people using Copilot for office work.
2.2 Memory
As well as custom instructions, Copilot can now remember things about you across chats and use your previous chats to shape its answers. These are the Saved memories and Chat history toggles in the same Personalisation settings, and both are switched on by default. This is convenient, but it has the same problem as custom instructions that you never update: if you told it in October that you were anxious about coding and had never written a line of R, it may still be treating you that way in March, and unlike custom instructions you won’t be able to see why. Click Manage saved memories now and again, delete anything that is out of date, and if you would rather it started from scratch each time, switch both toggles off. We’ll come back to why this matters in the “Be critical” section.
2.3 The Study and Learn agent
Copilot also has a built-in Study and Learn agent (in the left-hand sidebar under Agents) which has an Understand mode for exactly the kind of “explain this concept to me” questions this chapter is about, as well as Flashcards, Quiz, Matching and Fill in the blanks modes which we’ll use in the next chapter. It’s set up by Microsoft to behave a bit more like a tutor than the main chat does, and it will ask whether you want to upload your notes or slides as the source (see the copyright section in chapter 1 before you do).
You can use it for any of the activities in this chapter. So why are we bothering with custom instructions and prompts at all? Three reasons. First, the agent doesn’t know who you are, what course you’re on, or what you’ve been taught unless you tell it, and custom instructions are how you tell it once rather than every time. Second, you won’t always be using Copilot. If you use a different AI at some point, or Microsoft renames or removes the agent (which they will), the prompting skills transfer and the button doesn’t. Third, and most importantly, knowing how to tell an AI to behave like a tutor rather than an answer machine is the skill. The agent is Microsoft’s version of that; you should be able to build your own.
2.4 Asking questions
Now that you’ve got your tutor set up, you can ask it questions. There are two ways to do this and you should use both.
The first is to ask for an explanation:
- Why do I have to learn to code?
- What is the difference between short-term memory and working memory?
- Give me examples of between-subject designs
- What is the difference between a function and an argument?
- Rewrite this explanation in 100 words or less.
- Explain what each part of this code is doing:
ggplot(survey_data, aes(x = wait_time, y = call_time)) + geom_point() - Give me examples of when I would use different joins in R
The second is to ask it to make you work for the answer. Your custom instructions should already tell it not to give you the answer straight away, and the Study and Learn agent behaves a bit more like this by default, but for any individual question you can push it further by asking it to act as a Socratic tutor: instead of explaining, it asks you guiding questions so that you recall or work out the idea for yourself.
- “Act as a Socratic tutor. Do not tell me the answer yet. Ask me questions to guide me towards understanding correlation vs causation.”
- “Ask me a series of small questions until I can explain what a tibble is.”
- “Quiz me step by step on how the function
filter()works in R. Only give me the next hint if I get stuck.”
An explanation feels easier and the Socratic version feels slower, and that is the point. Reading an explanation is the AI equivalent of re-watching the lecture; having to produce the answer yourself is what makes it stick (the “Be critical” section below has the evidence).
Ask the AI three questions based on your course materials for this week. If the output doesn’t seem at the right level for you, edit your custom instructions and re-run the questions to see how the output changes.
For one of the questions, ask for two versions of the same answer (e.g., one simple, one more technical) and compare how the content changes.
For another, follow up by asking it to check your understanding as a Socratic tutor, and notice how different it feels to be asked rather than told.
2.5 Be critical
A personalised tutor is most powerful when it amplifies the cognitive processes that drive learning rather than replacing them. Three principles are central:
- Metacognition: monitor what you do and do not understand, then adjust your strategy and difficulty accordingly.
- Desirable difficulties: small, well-scaffolded challenges that require effort improve long-term retention when paired with feedback.
- Self-explanation: articulating why an answer is right, how a step works, or why an alternative is wrong deepens understanding and transfer.
Use your AI tutor to create space for retrieval, explanation, and calibration, not to short-circuit them. Based on the three principles, here are some things to watch out for.
If your custom instructions describe you as a nervous beginner (“I have never programmed before, keep it simple”), the AI will keep it simple long after you have stopped needing it to. Support that helps a novice gets in the way once you know what you’re doing, which is the expertise reversal effect. Good teaching fades the scaffolding as you improve; the AI won’t do that unless you tell it to. If you do not revise your instructions, or clear out Copilot’s saved memories, you “lock in” a static version of yourself (always a beginner), preventing growth and calibration.
Asking the AI questions like “Why do I have to learn to code?” produces an explanation, but does not force you to generate an answer yourself. Research on retrieval practice shows that passive review is less effective for long-term retention. Reading an explanation is the AI equivalent of re-watching the lecture.
You may accept AI answers at face value, especially if they “sound fluent.” The fluency illusion (e.g., when text feels easy to read) can lead to overconfidence, and people are reliably bad judges of their own learning. Unlike a lecturer who has genuine expertise, the AI cannot reliably detect specific misconceptions unless prompted very carefully. You might get partial reinforcement for incorrect ideas.
Learning research shows benefits when you try, fail, and then see the solution. If AI always provides a clean solution first, you miss the benefits of “desirable difficulties”. Related, over time, reliance on AI rather than working through the problem yourself will reduce your resilience, autonomy, and competence. You can only learn, and believe, that you are capable of difficult things if you try and succeed at doing difficult things. In Bandura’s terms these are mastery experiences, the strongest source of self-efficacy, and offloading the hard thing also offloads the evidence that you could have done it.
Tell Copilot who you are, what course you’re on, and your level of confidence, and tell it how to behave (no flattery, no answers before you’ve tried). This makes answers more tailored and less sycophantic. Update it as you improve.
Follow up, ask for comparisons, or request simpler/more technical versions. This is how you refine understanding.
Instead of answers, ask Copilot to quiz you step by step. This supports retrieval and deeper learning.
AI can be over-confident and sometimes wrong. Always compare to course materials and test your own understanding.
Reading an AI explanation is easier than doing the work yourself, but retrieval, practice, and self-explanation build stronger memory.