Build a Study Assistant Without Overspending

Choose one study task for your first version

A personal study assistant is easier to build when it starts with one clear job. Rather than asking it to manage an entire semester, choose something you repeat: explaining a difficult paragraph, turning lecture notes into practice questions, or checking your understanding of a concept.

For this walkthrough, build an assistant that accepts a short set of notes, explains one topic, and asks a follow-up question. Keep the first interface simple: a notes area, a question field, and a conversation view. You can add document uploads and saved sessions after this loop works.

Connect the chat API through a private backend

Your interface should send questions to a small backend that makes the chat API request. Keep the API key there, rather than exposing it in browser code. The backend can also enforce input limits and prevent repeated clicks from creating duplicate requests.

MyRelay gives an individual developer one key and one budget across its model menu, with request-level token and cost visibility. Create a MyRelay account, generate a key, and connect your backend using its API base address. If you already use an OpenAI-compatible SDK, configure its connection settings for the relay.

Start with one model, such as Gemini 2.5 Flash, and evaluate it on your own notes before adding alternatives. Keeping the model fixed makes it easier to see whether changes to your instructions actually improve the assistant.

Define how the assistant should teach

Write a reusable instruction that describes the teaching behavior you want. Ask the assistant to explain ideas in plain English, use the supplied notes as its main reference, and distinguish information found in those notes from additional background knowledge.

Include instructions to admit when the notes do not support an answer and to ask a clarifying question when necessary. Treat pasted material as reference content, not as instructions that can override the assistant's role. For homework practice, request a hint before a full solution so the tool supports your thinking rather than replacing it.

Send only the material needed for each question

For the first version, let the student paste the relevant section of notes manually. Give each section a readable heading, such as a lecture title or chapter name, and ask the assistant to identify which section supports its explanation. A reference is only useful if you can check it.

Avoid sending an entire textbook or every previous conversation with each question. Instead, include the current question, the relevant notes, and a small amount of recent discussion. If the project later grows into a searchable notebook, retrieve a few relevant passages before requesting an answer. You do not need that extra machinery to validate the basic idea.

Turn explanations into a short practice loop

After an explanation, have the assistant ask one question and wait for your answer. On the next turn, send your response alongside the practice question and supporting notes. Ask for specific feedback: what was correct, what was missing, and which idea deserves another attempt.

Keep a simple record of topics you struggled with and revisit them in later sessions. Save a brief topic summary rather than relying on an indefinitely growing chat history. Review generated answers against your course material, especially when a confident explanation introduces a claim you cannot locate.

Set spending limits before extending the project

Choose a personal weekly spending ceiling and track usage against it. Add an application-level request limit, cap response length, and restrict how much source material each request accepts. Make longer explanations an explicit choice rather than the default.

Use MyRelay's per-request token and cost records to inspect a typical session. Look for repeated context, unnecessarily long answers, and accidental retries. Test with a small set of familiar questions before changing models. Also avoid submitting classmates' personal information or restricted course materials without permission. The first milestone is modest: useful explanations, checkable feedback, and a session cost you understand.