Cleaning lecture notes with AI almost always introduces a quiet bug: the model fills the gaps with plausible-sounding facts the lecturer never said, and three weeks later you have memorized something the exam will not reward. The fix is not a smarter model, it is a stricter prompt. Every template below forces the model to rebuild structure, flag missing pieces with [GAP: ...], label every suggestion as “likely” or “needs verification”, and tie each self-test question back to the source line in your notes.
TL;DR: Paste your raw notes, run prompt 1 (structure), then prompts 2 and 4 (gaps + self-test). Keep the [GAP] markers visible so a hallucination can never sit silently next to a real note. For long audio transcripts, paste into a 1M-token model (Gemini 3.1 Pro, Claude Sonnet 4.6) rather than ChatGPT Plus, whose in-app context tops out around 320 pages. Then send the cleaned notes to the flashcard prompts to build spaced-repetition cards.
Which model handles your notes
The bottleneck is rarely reasoning, it is context length: a full semester of typed notes or a 90-minute transcript can blow past what a chat tool keeps in working memory. As of June 2026:
| Tool | In-app context | Best for | Notes |
|---|---|---|---|
| ChatGPT Plus ($20/mo) | ~320 pages (full 1M only on $200 Pro) | Short note batches, per-section cleanup | GPT-5.5 default; free tier caps context near 8K tokens |
| Claude Pro ($20/mo) | ~1M tokens (Sonnet 4.6) | Whole-course notes, long transcripts | Long context standard, not a paid add-on |
| Gemini 3.1 Pro (Google AI Pro, $19.99/mo) | 1M tokens (~1,500 A4 pages) | Multi-lecture sets, slides + audio | Multimodal: handles audio and image notes directly |
| NotebookLM (free) | 50 sources/notebook, 500K words each | Source-grounded cleanup, “cite the line” | 3 Audio Overviews/day, 50 chats/day on free |
If you want answers that never drift from your own material, NotebookLM is the safest base because it refuses to answer outside the sources you upload. For free-form rewriting and example generation, a 1M-token chat model is faster. See the study notes cleanup walkthrough for the end-to-end flow.
Best for
- Post-lecture cleanup the same evening, while the lecturer’s framing is fresh
- Exam revision when notes are too messy to review
- Building course notes from scratch out of slides plus shorthand
- Teaching prep and sharing readable notes with peers
1. Structure rebuilder
Below are my raw lecture notes from [course / topic]. Rebuild the structure with
headings, sub-headings, and bullet lists. Do NOT add facts I did not write; flag any
gap with [GAP: what is missing]. Keep my original wording where it is clear.
[paste notes]2. Gap-filling pass
Below are my structured lecture notes with [GAP: x] markers. Suggest what should fill
each gap based ONLY on the surrounding context. Label each suggestion "likely" or
"needs verification". Do not remove the [GAP] marker; append your suggestion after it.
[paste]3. Example generator per section
For each section in these notes, add 1 worked example. Use the formula or concept
exactly as stated in the notes. Tag each example [FROM NOTES] or [MY ADDITION] so I can
spot anything you invented.
[paste]4. Self-test question generator
Generate 3 self-test questions per section: (a) recall, (b) understanding,
(c) application. Format each as: question / answer / the exact source line in the notes.
If a question cannot be traced to a source line, drop it.
[paste]5. Glossary builder
Build a glossary of every technical term in these notes. For each term: a one-sentence
definition, the section where it first appears, and one example of correct usage drawn
from the notes.
[paste]6. “Confusing line” decoder
The following lines in my notes are shorthand or unclear. Expand each into a full
sentence. Flag any expansion that is not directly justified by the surrounding text
with [INFERRED].
[paste confusing lines]7. Cross-topic linker
Below are notes from 3 lectures: [lecture A / B / C]. Identify 5 cross-topic
connections the lectures imply but never state outright. Add a "Connections" section
and cite the two source lines each link bridges.
[paste]8. Highlight extractor
From these notes, extract the 10 most important points (formulas, definitions, claims).
Format: priority (must-know / should-know / nice-to-know) / one-line statement / the
page or line reference.
[paste]9. Audio-to-notes cleanup
Below is a raw transcript from a lecture recording. Clean it into structured notes:
remove filler words, fix grammar, and organize by topic. Preserve verbatim any moment
the lecturer flagged as "important", "on the exam", or "this will come up again".
[paste transcript]10. Diagram-from-notes prompt
Below is a section of notes describing a process or system. Describe a diagram that
captures it. Output: nodes, edges, labels, and the 1 question the diagram should help
me answer. Do not add steps that are not in the notes.
[paste section]11. Personal-study summary
Convert these notes into a 1-page personal study summary. Focus on what I struggle with
most: [topics]. Use my own vocabulary from the notes. Add 5 self-test questions tied to
my weak areas, each with the source line.
[paste]Model-version notes (June 2026)
- Placeholders: swap every
[bracketed]field for your own text before sending. The brackets are just markers, not syntax. - Long transcripts: GPT-5.5 in ChatGPT Plus keeps roughly 320 pages in working memory; a 90-minute transcript can exceed that. Sonnet 4.6 and Gemini 3.1 Pro both hold ~1M tokens (~1,500 pages), so a full module fits in one pass.
- Source-grounded option: if you cannot risk a single invented fact, upload your notes to NotebookLM (free) and run prompts 4, 5, and 8 there; it answers only from your uploads and links each claim back to the source.
- Reasoning modes: for gap-filling (prompt 2) and cross-topic links (prompt 7), pick the “Thinking” mode in ChatGPT or Gemini so the model checks itself before guessing.
FAQ
Will the AI still invent facts even with these prompts?
Less often, but not never. The [GAP] and “needs verification” labels are designed so that when it does guess, the guess is visible instead of blended into your real notes. Always spot-check anything tagged “likely” or [INFERRED] against the slides.
ChatGPT, Claude, or Gemini for messy lecture notes? For a single lecture, any of them work. For a whole semester or a long transcript, use Claude Sonnet 4.6 or Gemini 3.1 Pro because their ~1M-token context holds the full set without dropping earlier sections. ChatGPT Plus is fine for per-section cleanup but truncates around 320 pages in-app.
Is NotebookLM better than a chat model for this? For accuracy, yes: it refuses to answer outside your uploaded sources and cites the exact line, which is ideal for prompts 4, 5, and 8. For free-form rewriting and example generation (prompts 3 and 11), a chat model is faster and more flexible. Many students use both.
How do I clean up an audio recording of a lecture? Transcribe first (Gemini 3.1 Pro and NotebookLM both accept audio directly), then run prompt 9. The key is preserving every “this will be on the exam” callout verbatim, which generic cleanup tends to delete.
Can I do this on a free plan? Yes. NotebookLM’s free tier gives 50 sources per notebook (up to 500K words each) and 50 chat queries a day, which covers a typical course. Free ChatGPT works for short notes but caps context near 8K tokens, so paste one section at a time.
Common mistakes
- Adding facts the lecturer never said. The exam ranks the lecturer’s framing, not the model’s.
- Deleting surrounding context you will need to understand a line later in the term.
- Dropping the “needs verification” flag, so confident hallucinations sit beside real notes.
- Skipping the self-test step (prompt 4) that turns notes into actual revision material.
- One giant block of text with no headings, which makes spaced review impossible.
- Cleaning an audio transcript without keeping the lecturer’s “this will be on the exam” callouts.
Related
- Theory simplification prompts
- Concept comparison prompts
- Quiz generation prompts
- Flashcard prompts
- Study plan prompts
- Generate Follow-Up Questions for a Lecture With AI
- Clean Up Lecture Notes With AI: From Messy Scratch to Revision-Ready
- How to Use AI to Clean Up Lecture Notes: From Raw Capture to a Reviewable Outline