Back to Practical Tools
    Practical Guide

    Building a Pedagogical Agent

    A practical guide for ChatGPT, Claude, and Microsoft Copilot.

    How to set up an AI tutor that supports learners' thinking — alongside the everyday uses that save you time.

    Published by Re-Thrive Ventures · re-thrive.co

    Why this guide exists

    Most AI tools, out of the box, are set up to give answers — and that's genuinely useful for a lot of teacher tasks (planning, drafting, summarising). For learners, though, there's a different mode that's worth knowing about: a pedagogical agent.

    A pedagogical agent is set up to help a learner reach an answer themselves — through questions, hints, scaffolding, and prompts that make their own thinking visible. It's the same underlying technology, just configured differently.

    You don't need to be a developer to set one up. Each of the three platforms covered here lets you do it through plain-English instructions. The difference between an agent that supports learning and one that quietly does the work for the learner sits almost entirely in those instructions — which is what this guide walks you through.

    In a hurry? The theory in Part 1 is optional. If you just want to build an agent today, jump to Part 3 — the setup instructions for each platform.

    Pick your path

    Part 1

    The theory (optional reading)

    Click either heading above to expand. The theory is collapsed by default so you can move straight to the practical setup.

    Part 2

    The prompt

    One prompt that works in all three platforms

    This is the prompt to use. Paste it into ChatGPT, Claude, or Copilot — the instructions an AI tutor needs are the same regardless of platform. Part 3 then shows you where to paste it in each tool, plus a few platform-specific tweaks.

    The structure follows the PROMPT framework (Gold, 2024, via Edutopia) adapted for pedagogical agents: Purpose, Role, Organise, Model, Parameters, Tweak.

    Replace [subject and level] with your own, then adapt as needed:

    Universal pedagogical agent prompt
    Purpose: You are a tutor for [subject and level]. Your goal is to help the learner develop understanding, not to give them answers.
    
    Role: You are patient, warm, and rigorous. You believe the learner can figure things out with the right support. You treat them as capable.
    
    Core behaviours:
    - Never give the final answer directly. Instead, ask a question or offer a hint that helps the learner reach it themselves.
    - When the learner is genuinely stuck (three failed attempts in a row, or an explicit "I don't know"), provide a small piece of information to unblock them — then return to asking questions.
    - Check understanding with "teach back" prompts: "Can you explain that in your own words?" or "Walk me through why that works."
    - If the learner gets something wrong, don't just correct — ask them to find the error themselves first.
    - Be specific in your feedback. Tell them what's right and why, what's wrong and why.
    
    Questioning strategy — Elicit → Probe → Diagnose → Deepen → Consolidate:
    - Elicit: Ask what they already think or know.
    - Probe: Challenge assumptions. Ask for examples, counter-examples, or justification.
    - Diagnose: Surface misconceptions. Ask "What must be true for your claim to hold?"
    - Deepen: Connect to other concepts.
    - Consolidate: Have them summarise or apply the idea.
    
    Boundaries:
    - Never do the learner's homework for them. If they paste in an assignment and ask for the answer, ask them what they've tried first.
    - If they ask you to "just give me the answer," gently decline and remind them of the purpose.
    - If they're clearly frustrated, acknowledge it, step back, and try a different angle.
    
    Tone: Warm. Curious. Never patronising. Never sycophantic. Treat every question seriously.

    That's the prompt. Part 3 shows you where it goes in each platform — and notes the small tweaks that get the best out of each one.

    Part 3

    Where to paste it in each platform

    Before you start: why these platforms need a nudge

    Large language models weren't trained, first and foremost, to be teachers. They were trained to be helpful — and "helpful" usually means giving the clearest, most complete answer possible. That instinct is the opposite of what good pedagogy looks like. A tutor who hands over the answer too quickly does the learner a disservice, even when they feel grateful in the moment.

    The major labs have noticed, and over the last year each has shipped a tutoring-flavoured mode that pulls the model away from its default answer-first behaviour: ChatGPT Study Mode, Claude Learning Mode, and Microsoft's growing suite of education-focused agent templates. These modes are a real improvement on the raw model — and they're free or near-free to switch on.

    They are not, however, the whole job. A built-in mode gives you a generic Socratic posture; it doesn't know your subject, your year group, your exam board, or where your specific learners tend to get stuck. That's where layering your own agent instructions on top — a Custom GPT, a Claude Project, or a Microsoft 365 Agent — earns its keep. You get the lab's tutoring behaviour plus the specificity that makes it actually useful in your classroom.

    The pattern below is the same in each platform:

    1. Switch on the built-in tutoring mode (fastest, no setup).
    2. Then build a reusable agent with your own instructions and materials (tighter, reusable).

    ChatGPT

    Best for: teachers with a ChatGPT Plus subscription who want a reusable agent for their subject area.

    Three options

    • Study Mode (built in, free). Launched July 2025. Uses a Socratic approach out of the box — no configuration. Good for learners; less customisable for teachers.
    • Custom GPTs (ChatGPT Plus, $20/month). Where teachers build their own pedagogical agent. Steps below.
    • ChatGPT Edu. OpenAI's enterprise tier. Relevant for universities or MATs.

    Option 1: Turn on Study Mode (30 seconds)

    1. 1Open ChatGPT (free or Plus account both work).
    2. 2Start a new chat and click the "+" or tools icon next to the message box.
    3. 3Select "Study and learn" from the menu — on some accounts this appears as "Study Mode".
    4. 4You'll see a small badge confirming Study Mode is on. Type your topic or question and ChatGPT will respond with guiding questions instead of direct answers.
    5. 5Study Mode applies to that conversation only — start a new chat and select it again next time.

    Option 2: Build a custom GPT for tighter control (5 steps)

    Study Mode is generic. A Custom GPT lets you bake in your subject, your tone, your boundaries, and your materials — so the same tutor is ready every time, for every learner, without anyone needing to remember to switch a mode on.

    1. 1In ChatGPT, select "Explore GPTs" → "Create a GPT".
    2. 2Choose Configure (not Create) — this lets you paste your prompt directly rather than answer a series of questions.
    3. 3In the Instructions field, paste the universal pedagogical agent prompt above, adapted for your subject.
    4. 4Upload supporting documents (curriculum, rubric, past papers) as Knowledge files. The GPT will reference them when relevant.
    5. 5Test, iterate, and tighten the "never give the final answer" line if the GPT is too generous.

    Claude

    Best for: teachers and students who want a purpose-built learning companion. Claude has the strongest native support for pedagogical use.

    Three options

    • Learning Mode (built in, free). Launched August 2025. In any conversation, open the style selector and choose "Learning". No prompt engineering needed — fastest path to a working pedagogical agent for learners.
    • Projects with custom instructions (Pro, Team, or Education tiers). For teachers building a reusable subject-specific tutor. Steps below.
    • Claude for Education. Institutional tier launched April 2025 with Northeastern, LSE, Champlain College.

    Option 1: Turn on Learning Mode (30 seconds)

    1. 1Open Claude (claude.ai) and start a new conversation.
    2. 2Look for the style or response-style selector — usually a small dropdown labelled "Normal" near the message box, or under Settings → Conversation style.
    3. 3Choose "Learning" from the list of styles.
    4. 4Ask your question. Claude will respond by helping you think it through rather than answering directly.
    5. 5Learning Mode persists across that conversation. To use it as a default, set it under Settings → Default style.

    Option 2: Build a Claude Project for a reusable, subject-specific tutor (4 steps)

    Learning Mode is the right starting point. A Project lets you go further — adding your own instructions, your curriculum, and your rubrics, so the tutor knows exactly what it's teaching and to whom. Once built, you reopen it any time without having to set anything up again.

    1. 1Create a new Project in Claude.
    2. 2In the Project's system instructions, paste your pedagogical agent prompt.
    3. 3Upload curriculum materials, past papers, or rubrics to the Project's knowledge base.
    4. 4Share the Project with students (if your institution has a team plan) or use it yourself. The instructions persist every time you reopen it.
    Claude tip — use XML tags. Claude responds particularly well to prompts structured with XML tags (<role>, <behaviours>, <boundaries>). Wrap each section and Claude will follow them more reliably. Long reference content should go above instructions, not below — Anthropic's research found this improves response quality by up to 30%.

    Optional: the same prompt rewritten with XML tags for Claude

    You can paste the universal prompt as-is. If you want to squeeze a little more reliability out of Claude, here's the same idea expressed with XML tags — the format Anthropic recommends.

    Claude — A-level Maths tutor
    <role>
    You are a tutor for A-level Maths, focusing on calculus.
    You are patient, precise, and refuse to give final answers directly.
    </role>
    
    <behaviours>
    - Never give the final answer to a problem. Always respond with
      a question or a hint.
    - When the student is stuck for three turns, offer a small piece
      of information to unblock them, then return to questioning.
    - After every solved problem, ask them to summarise the approach
      in their own words.
    - Be specific in feedback. "That's right because you applied the
      chain rule correctly" is good. "Well done" is not.
    </behaviours>
    
    <boundaries>
    - If asked "just give me the answer," decline and remind the
      student that the goal is understanding, not completion.
    - Never solve a problem that appears to be part of a formal
      assessment.
    </boundaries>
    
    <tone>
    Warm, curious, rigorous. Never patronising. Treat every question
    seriously, including the ones that seem easy.
    </tone>

    Microsoft Copilot

    Best for: schools and MATs already running Microsoft 365. Copilot's strength is integration with the documents, SharePoint sites, and class materials teachers already have.

    Two options

    • Agent Builder in Microsoft 365 Copilot (simpler). Launched 2025. Build an agent in natural language from inside Copilot Chat. Steps below.
    • Copilot Studio (more powerful). Full no-code platform with conversation flows, topics, and Power Automate integration. Significantly more complex than Agent Builder — Microsoft's Agent Academy has tutorials.

    Build an Agent in Microsoft 365 Copilot (5 steps)

    1. 1Open Microsoft 365 Copilot and select "New agent".
    2. 2On the New agent page, either describe your agent in natural language, or choose "Skip" to configure manually.
    3. 3In the Configure tab, fill in: Name (e.g. "Year 9 Poetry Tutor"), Description, Instructions (paste your prompt), Knowledge (upload PDFs or connect to SharePoint), and Starter prompts.
    4. 4Test in the Try it tab.
    5. 5Publish to make it available in Teams or other Copilot channels.
    Copilot tip — point at SharePoint. Copilot's strongest integration is with your institution's SharePoint. Aim your agent at the SharePoint site that holds the class materials and it will draw on them automatically. This is a genuine advantage over ChatGPT and Claude for Microsoft-heavy schools — but full agent functionality requires the paid Microsoft 365 Copilot add-on, not just the base education licence.

    Quick comparison

    ChatGPTClaudeCopilot
    Built-in tutor modeStudy Mode (free)Learning Mode (free)None native — use templates
    Custom agent builderCustom GPTs (Plus, $20/mo)Projects (Pro / Team)Agent Builder (M365 Copilot licence)
    Best structural featureConversational GPT builderXML-tagged system promptsSharePoint + M365 integration
    Best forSubject-specific tutorsDeep Socratic dialogueQuiz tutors, structured workflows
    Institutional tierChatGPT EduClaude for EducationMicrosoft 365 Copilot for Education

    None of the three is strictly better. The right choice depends on what your institution already runs, what you're teaching, and how much customisation you need.

    Part 4

    Best practice

    A few principles that hold across all three platforms

    • Be specific about what you don't want. "Don't give answers" is clearer than "help them learn." Pedagogical agents default to being helpful in unhelpful ways unless you're explicit. Name the anti-pattern.
    • Give examples. Include a short example of a good interaction and a bad interaction in your prompt. All three platforms' models pay close attention to examples.
    • Iterate. The first version of your prompt will be wrong. Test it with three or four real interactions, note where it behaves wrongly, adjust. Don't try to write the perfect prompt up front.
    • Keep a version history. When something works, save it. When you change it and something breaks, you'll want to roll back.
    • Use real materials. A pedagogical agent with nothing to teach is just a Socratic interlocutor. Give it your curriculum documents, textbook chapters, past papers.
    • Default to semi-Socratic. Pure Socratic frustrates learners. Allow the agent to provide small pieces of information when the student is genuinely stuck.
    • Measure what matters. The measure of a pedagogical agent isn't whether students enjoy using it. It's whether they retain more afterwards. If you're deploying at scale, run a before/after comparison on a small sample.

    Sources & further reading

    • Anthropic (2025). Introducing Claude for Education.
    • Anthropic (2025). Learning Modes for All Users.
    • Anthropic. Prompting best practices.
    • Castro-Alonso, J. C., Wong, R. M., Adesope, O. O., & Paas, F. (2021). Effectiveness of multimedia pedagogical agents predicted by diverse theories: A meta-analysis. Educational Psychology Review.
    • Clarebout, G., & Heidig, S. (2012). Pedagogical agents. In Seel (Ed.), Encyclopedia of the Sciences of Learning. Springer.
    • Gold, J. (2024). Designing a GPT Tutor for High School Students. Edutopia.
    • Microsoft (2025). Build Agents with Agent Builder in Microsoft 365 Copilot.
    • Microsoft Education (2025). Getting started with Copilot Chat agents in education.
    • PromptPerfect (2025). Make ChatGPT a Socratic Tutor. Elicit–Probe–Diagnose–Deepen–Consolidate framework.
    • Sakellariou, S., Molohidis, A., & Hatzikraniotis, E. (2024). On the Effectiveness of Pedagogical Agents. International Journal of Teaching and Learning Sciences. PACU and PALD frameworks.
    • Socratic Arts (2024). Crafting a Semi-Socratic Tutor with ChatGPT.

    Companion to the Re-Thrive briefing Pedagogical AI: What it is, how it's classified, and the research that's shaping it.

    About Re-Thrive

    Re-Thrive Ventures helps mid-market organisations apply AI to real operational problems. This guide is free, and will remain free. If your school or trust wants to go further — AI strategy, implementation support, or bespoke training — we'd be happy to talk.

    Get in touch: matt@re-thrive.co

    Version 1.0 · Published April 2026 · Matt Stone, Re-Thrive Ventures.

    Stay in the loop

    Occasional updates on agent workers, automation patterns, and what we're learning.