Which Model Should Summarise Your Lecture Notes?

Start with one lecture and a clear study outcome

The best model for summarising lecture notes is the least expensive option that reliably preserves what you need to learn. A polished paragraph is not enough if it drops a condition from a theorem or turns a tentative explanation into a fact.

Begin with one representative lecture rather than uploading a whole semester. Decide whether you want a revision sheet, an explanation of difficult concepts or a list of questions to practise. For a first test, aim for a concise revision sheet with definitions, main arguments, worked examples and unresolved questions.

Prepare notes that a model can interpret honestly

Remove repeated headers, attendance details and unrelated messages. Keep the lecture title, section headings and any information that explains the order of the material. Do not remove examples just because they look repetitive; they may show when a rule applies.

Check copied equations and tables before submitting them. If your notes contain phrases such as “this increases” without naming the quantity, mark the ambiguity. A model cannot reliably recover a missing diagram or a lecturer’s unrecorded explanation. Ask it to flag gaps rather than fill them with plausible guesses.

Try Gemini 2.5 Flash as your first candidate

Gemini 2.5 Flash is a reasonable starting candidate for a straightforward summarisation task. Treat that as a hypothesis to test, not a guarantee that it will be the cheapest or most accurate choice for your notes. Check the current rate and your actual request cost before committing to repeated use.

Through MyRelay, you can use one key and one budget across supported providers, then inspect each request’s tokens and cost. Create a MyRelay account to run a small trial and see what your own lecture notes cost to process.

Give the first summary a precise brief

Tell the model who the summary is for and how much detail to retain. A useful instruction is: “Summarise these lecture notes for a student revising this topic. Use only the supplied material. Preserve definitions, assumptions, exceptions and important examples. Flag unclear passages and separate unanswered questions from established points.”

Add an output limit, such as approximately 600 words, and request headings that follow the lecture’s structure. If you want practice questions, ask for a small number with answers supported by the notes. Keeping the requested output bounded helps control both reading time and output-token spending.

Check the result against the difficult passages

Read the summary beside the original notes. Start with the parts where an error would matter most: definitions, numerical values, causal claims, equations and exceptions. Do not judge quality only by how smoothly the answer reads.

• Check that every major section appears and that the summary preserves the lecturer’s distinctions.
• Trace important claims back to the notes, especially statements that sound more certain than the source.
• Confirm that practice answers are supported and that missing information is acknowledged.

If the problems come from unclear instructions, revise the brief before changing models. One targeted correction is more informative than repeatedly requesting a “better” summary.

Compare Claude Sonnet 5.5 only when needed

If the first candidate repeatedly loses important relationships or mishandles dense reasoning, test Claude Sonnet 5.5 on the same lecture with the same instructions. This is a comparison, not an automatic upgrade. Evaluate whether it actually improves the errors you identified.

Compare both summaries against the source and inspect the cost of each request. A higher-cost result is worthwhile only if its improvement matters to your study task. If both pass your checks, keep the lower-cost option for similar lectures.

Reuse the tested workflow without resending everything

Save the instructions that worked and process new lectures individually. For long notes, divide them at topic boundaries, retaining the definitions each section needs. Combine the section summaries afterward, checking for lost connections.

Review costs periodically rather than assuming every lecture will cost the same. Longer inputs, longer outputs and repeated attempts all affect spending. Keep the original notes as your reference: the summary should make revision easier, not become an unchecked replacement for the course material.