萃靈 Essevin
It is 9 p.m. A child points at a word problem and says, “I don’t understand.” The parent can see roughly how to calculate the answer but cannot find the first sentence that will help. The easiest response is to send a photo to AI and request the answer. The number appears quickly; the child’s misunderstanding remains hidden.
Using AI as a tutor should not mean letting it do the homework. Make it diagnose first, reveal hints gradually, and verify the reasoning at the end. The parent stays involved to observe how the child thinks and to check that the question and method are appropriate.
Use AI as a coach, not an answer sheet
A good tutor does not write the full calculation as soon as the student sits down. They ask: What information do we have? What are we trying to find? What have you tried? An AI model can follow the same rhythm, but the prompt must require it.
It is like helping a child learn to ride a bicycle. Holding the back of the seat helps the child balance; it should not mean pedalling the entire route for them. Too little support stops progress. Too much support hides whether balance was learned.
Give four pieces of context
School year or learning level. “Primary 3” is more useful than age alone.
The complete question. Type it accurately or upload a clear image with units, diagrams, and options visible.
The child’s attempt. Include a wrong calculation; it shows where the reasoning changed direction.
The goal. Decide whether the session is for understanding a concept, completing one question, or practising the same structure.
Do not provide the child’s name, school, class, student number, or any personal information that the maths task does not need. Crop or cover it before uploading an image.
A four-stage tutoring flow
Stage 1: Ask one diagnostic question
Make the model ask one question and wait. The first question should usually test understanding of the problem, not arithmetic:
Do not solve the problem yet. Ask the child one question to check what information the problem gives and what it asks us to find. Wait for me to provide the child’s answer before continuing.If the model asks five questions at once, a child may answer only the final one. One-at-a-time questioning exposes the reasoning process.
Stage 2: Reveal hints in three levels
When the child is stuck, do not jump to a complete calculation. Ask for three levels and reveal only one at a time:
Hint 1: point to the important word or relationship;
Hint 2: demonstrate the same relationship with smaller numbers or objects;
Hint 3: provide a calculation frame while leaving the important step to the child.
For equal sharing, the first hint might ask whether every group needs the same amount. The second can use twelve sweets shared by three people. The third can introduce the frame total ÷ number of groups = amount per group.
Stage 3: Ask the child to explain why
After getting a number, ask the child to explain in one sentence why addition, subtraction, multiplication, or division was appropriate. Matching the expected answer does not by itself prove understanding.
Stage 4: Verify independently and vary the problem
Ask for another check: substitute the answer into the question, use the inverse operation, or draw a simple quantity relationship. Then generate one new problem with the same structure and different numbers. Do not reveal its answer until the child attempts it.
A reusable parent prompt
You are a patient primary maths tutor working with a <school year> student.
Full question: <paste the question>
The child’s attempt: <paste the calculation, answer, or “not started”>
Goal: <understand the concept / complete this problem / practise this type>
Rules:
1. Check that the question is complete. If text or a diagram is unclear, ask the parent; do not guess.
2. Do not give the answer or a complete calculation immediately.
3. Ask only one question at a time and wait for our response.
4. If the child is stuck, reveal one level at a time: key-word hint, smaller example, calculation frame.
5. Use language and methods suitable for the school year. Do not introduce concepts the child has not learned.
6. After an answer, ask why before identifying what is correct and the one next correction.
7. Finish with a different verification method and one similar practice problem. Do not provide its answer until the child attempts it.
8. Clearly flag uncertainty and recommend checking with the teacher when the question, answer, or teaching method remains unclear.What the parent should observe
A longer AI response is not necessarily better. Watch whether the child can restate the problem, explain the operation chosen, and solve the same structure with different numbers. Those signals reveal more than whether one answer happened to be correct.
If the model uses a method the class has not covered, add: “Use the method shown in this term’s . Do not use equations or another unlearned concept.” A completed textbook example may help, but remove names and school details first and ask the model to follow the steps, not copy an answer.
The core prompt and observation points are now ready to save. The next section turns one problem into a short practice sequence.
Build a ten-minute practice sequence
After one problem, do not request ten more. Ask the model to:
state in one sentence the concept that caused difficulty;
give one problem with different numbers but the same structure; and
give one problem in a different setting with the same quantity relationship.
Give one problem at a time and wait for the child’s attempt. This keeps practice diagnostic instead of turning it into another worksheet to rush through.
Common failure modes
The image is incomplete. A cropped diagram, unclear unit, or missing option may not stop a model from answering. Require it to ask when information is missing.
A correct answer is treated as a correct teaching method. The model may use a method from a later school year. The school’s materials and teacher remain the reference for the expected method.
The child only says “next”. If the parent enters every response, the model cannot tell whether the child understands. Ask the child to explain first, then transcribe the answer.
The upload contains unnecessary personal data. Maths does not require a name, school, class, or student number. Crop first.
AI becomes the only source of truth. If the wording is ambiguous, the result conflicts with the textbook, or the child remains confused, keep the reasoning steps and ask the teacher.
The child still completes the work
The value of AI here is not turning a five-minute problem into a five-second answer. It is breaking down the questions a parent may find difficult to phrase: diagnose, hint, explain, and verify. The session ends when the child can complete a structurally similar problem without a hint.
This method works in any general-purpose AI chat. Whichever model you use, remove unnecessary personal data, verify that the full question is present, and have a parent or teacher review the teaching content. Tool choice does not replace these steps.
Information in this article is current as of 12 August 2026 and is provided for general reference only; it does not constitute advice of any kind. Third-party product features, pricing and policies are subject to their official announcements. Essevin service details are as shown on essevin.com and in the console.