Build an AI Math Tutor From Your Problem Sets

TL;DR: Math tutors and teachers build AI math companions by uploading worked examples, solution walkthroughs, and hint sequences. The agent guides students through problems step by step — providing hints rather than answers, asking where they're stuck, and never jumping ahead of what the student has tried.

An AI math tutor configured from your course materials on Alysium — uploaded with your problem sets and curriculum — gives students a knowledgeable guide through your specific content.

The hardest part of math homework isn't the problem itself — it's knowing where you went wrong. A student who gets a wrong answer and has no one to help them trace the error either gives up or guesses at what they did differently, neither of which builds understanding. What they need is something that looks at where they are in the solution, identifies the step that went sideways, and asks the right question to help them see it.

An AI math tutor built from your course materials — not generic internet content — knows where students are in your curriculum and can scaffold from the concepts you've already covered.

That's a tutoring skill. And it's one that a well-configured AI math companion can approximate — not by giving the answer, but by working the Socratic hint sequence that a good tutor uses.

Step 1: Decide What Problem Types to Cover

Before uploading anything, define the scope: which problem types, which difficulty levels, which course topics. A calculus AI tutor covers integration techniques differently than a pre-algebra AI tutor. The more specific your scope, the more accurate your agent's guidance will be. A focused agent for limits and derivatives serves calculus students better than a broad "all math" agent that lacks depth in any single area.

Write a one-paragraph scope statement: "This agent helps students work through problems from [course name] covering [topic list]. It guides students through steps using hints and questions but does not provide final answers or complete worked solutions." That scope statement becomes the first section of your instruction set and sets the constraint that all subsequent behavior operates within.

Step 2: Build Your Worked Example Library

The knowledge base is your worked examples — ideally 3–5 fully solved problems per topic, showing each step explicitly with brief explanations of why that step is appropriate. The key is that the examples should reflect your teaching style and notation, not a generic textbook's. If you use a specific substitution method, your examples should show that method. If you use a particular way of laying out integration by parts, your examples should match what students see in class.

The distinction between a worked example and a solution key matters here. A solution key shows the final answer and the steps. A worked example shows the steps, the reasoning behind each step, and the common errors to watch for at each transition. Build your examples closer to the second format — the agent retrieves from this content when students ask "why do we do this step?", and a worked example with reasoning produces a far better answer than a solution key with steps.

A practical technique for ensuring worked examples produce useful retrieval: structure each example with explicit headings for each phase. 'Setup,' 'Approach,' 'Execution,' and 'Verification' headers within a worked example help the agent retrieve the right portion of the example when a student asks a specific process question. A student who asks 'how do I verify my answer?' gets the verification section, not the full walkthrough. That precision only happens when the example document is internally structured rather than presented as undifferentiated flowing text.

Step 3: Add a Hint Sequence Document

The single most valuable document in a math tutor knowledge base is a hint sequence: for each problem type, a series of graduated hints that guide a stuck student toward the solution without giving it away. A hint sequence for a related rates problem might look like: Hint 1: "Draw a diagram and label all the quantities that are changing." Hint 2: "Write an equation relating the quantities from your diagram." Hint 3: "Differentiate both sides with respect to time using the chain rule." Each hint moves the student forward without completing the step for them.

Build hint sequences for the 5–8 most commonly stuck points in your course — the places where students consistently hit a wall. These are the transitions that feel obvious to an expert but aren't to a novice: setting up the integral, choosing the substitution, simplifying before applying a rule. If you've taught the course for more than a semester, you know exactly where these walls are. That knowledge, encoded as hint sequences, is the highest-value content in your math tutor knowledge base.

Step 4: Write the Hint-First Instruction

The instruction set for a math tutor has one primary constraint: the agent provides hints before answers, and final answers only after the student has shown their work. A direct instruction: "When a student asks how to solve a problem, do not provide the solution. Instead, ask where they are in their attempt and what step is confusing them. Provide one hint at a time and wait for the student's response before providing the next hint. Only reveal a complete step after the student has attempted it and shown their work."

This instruction pattern keeps the interaction in tutoring mode rather than answer-retrieval mode. Students who work through a problem with hints and then check their result against the worked example learn more than students who read the solution. The instruction makes hint-first behavior the default for every interaction, not just when the student explicitly asks for a hint.

One additional instruction that improves hint quality significantly: 'When providing a hint, state the specific step where the student should focus their attention before explaining what to do.' A vague hint like 'check your algebra' gives a student nowhere to look. A specific hint like 'look at the step where you moved the variable from the right side to the left — is the sign correct?' tells the student exactly where to direct their attention. That specificity comes from the combination of the hint-first instruction and the error pattern document — the agent knows both to give hints and to be specific about which step.

Step 5: Configure Error Pattern Recognition

The most useful capability in a math tutor agent is recognizing common errors. If a student shows their work and it contains a common algebraic mistake — sign error, incorrect distribution, wrong derivative rule — the agent should be able to identify the error type and ask a question that helps the student catch it. This requires uploading an error pattern document: the 10–15 most common mistakes you see in student work, with the question you'd ask to help a student recognize each one.

An error pattern entry might look like: "Common error: forgetting the chain rule when differentiating composite functions. Recognition question: 'Is the function inside the square root changing with respect to x? If so, what rule applies when you differentiate a function of a function?'" That question plants the recogni­tion without stating the error — exactly what a good tutor would do.

The error pattern document works best when organized by the consequence of the error rather than its technical name. Students don't know they're making a 'chain rule omission error' — they know their derivative 'came out wrong.' Structure each entry as: 'When a student gets [type of wrong answer], the likely cause is [error], and the recognition question is [question].' That consequence-first organization helps the agent identify the error when a student describes their result rather than when they describe their process.

Step 6: Test With Problems Students Find Hard

Test the agent on the problems students historically struggle with most in your course — not the standard examples, but the ones that appear on exams and produce the widest grade spread. Ask the agent a problem and respond as a confused student would: giving a partial attempt, asking what to do next, making the common error for that problem type. Evaluate whether the hint sequence moves you toward the solution or whether the agent's responses are too vague to be actionable.

The specific test: give a partial incorrect solution and ask "what did I do wrong?" The agent should be able to identify the step where the work diverged from correct procedure and ask a question about that specific step. If it can't — if it responds with general advice like "check your algebra" — the error pattern document needs more specificity for that problem type.

One additional testing scenario that's often skipped: test the agent when the student gives up before completing any work. Type 'I have no idea where to start' for a problem type in your knowledge base. The agent should be able to provide a first-step hint that's specific enough to actually begin — not 'read the problem carefully' but 'draw a diagram and label all the quantities the problem gives you.' A stuck student at the start of a problem is the highest-stakes interaction the agent will have; testing this scenario specifically ensures the agent can handle it.

Step 7: Deploy and Integrate With Your Tutoring Flow

The deployment that works best for math tutors: position the AI companion as the first stop before a student requests a live tutoring session. "Before booking a session with me, try the AI companion for at least 15 minutes — if you're still stuck after that, I'll know you've already worked through the basics and we can start at the harder question." This positions the AI as a complement to live tutoring rather than a replacement, sets productive expectations, and ensures that live sessions focus on genuinely complex problems.

Build your math tutoring companion today. Start free on Alysium — upload your worked examples and hint sequences, configure your instructions, and deploy before your next assignment is due.

Private tutors who've deployed math companions report a specific workflow improvement: students who complete a homework set send their work to the agent first and arrive to sessions with a list of the specific steps they couldn't resolve. The tutoring session starts immediately at the genuine difficulty rather than spending 20 minutes establishing which problems were hard. That session efficiency improvement — starting at the hard problem rather than triaging — is worth the 60 minutes of initial agent build time in the first month of use alone.