TL;DR
Paste your raw notes into ChatGPT, Claude, or Gemini with the prompt below. Tell it to restructure and define, never invent — ambiguous spots get flagged as [UNCLEAR], not silently filled. The single biggest failure mode is hallucination: the model smoothing over your half-finished notes with a plausible explanation the professor never gave. The fix is one instruction (“do not resolve gaps”) plus a 30-second check that gaps were actually flagged. A 90-minute lecture should produce 2-5 gap markers; zero gaps usually means the model invented its way past your ambiguity.
The task
You took fast, messy notes during a 90-minute lecture. The page is full of abbreviations, half-finished sentences, and arrows pointing at things you no longer remember. You want a cleaned-up version you can read tomorrow that highlights what you do not yet understand. The trap with AI here is hallucination: it will happily smooth over an ambiguous note by inventing a plausible explanation that does not match what was said in class. In study material, that is worse than a typo — you will memorize the fabrication and walk into the exam confident about something wrong.
When AI helps — and when it does not
AI is genuinely good at three things with notes: restructuring fragments into a hierarchy, defining standard terminology, and surfacing internal inconsistencies. It is poor at filling in content that is not in your notes, and that is exactly where fabrication is most dangerous.
So split the job explicitly. Let the model reorganize and define; forbid it from resolving anything ambiguous. Tell it to mark ambiguity as a gap, not to guess at it. The goal is reviewable notes, not a polished textbook chapter.
Which model to use (as of June 2026)
Any current general-purpose model handles this. The differences that matter for note cleanup are context window (how much you can paste at once) and the free tier (because most students will not pay).
| Model / plan | Free? | Context you can paste | Notes for this task |
|---|---|---|---|
| ChatGPT Free (GPT-5.5) | Yes | Short (tight limits; US Free shows ads) | Fine for one lecture of text |
| ChatGPT Plus ($20/mo) | No | ~320 pages in-app | Comfortable for a full module |
| Claude Free (Sonnet 4.6) | Yes | Limited | Lowest hallucination in note cleanup |
| Claude Pro ($20/mo) | No | 1M tokens standard | Best for whole-term consolidation |
| Google AI Pro ($19.99/mo) | No (free Gemini tier exists) | 1M tokens (Gemini 3.1 Pro) | Strong on long PDFs and photos |
Practical takeaway: for a single lecture of pasted text, the free tier of any of the three is enough. If you are merging a whole course (dozens of lectures) into one document, you want a 1M-token context — Claude Opus 4.7 / Sonnet 4.6 and Gemini 3.1 Pro all offer 1M as standard (as of June 2026). Independent OCR and hallucination tests through early 2026 give Claude a reputation for lower hallucination on messy input, which is the property you care about most here. See our ChatGPT vs Claude vs Gemini comparison for the full breakdown.
What to feed the AI
- The raw notes — paste everything, including abbreviations and arrows
- Subject area (Org Chem, Constitutional Law, Linear Algebra)
- Lecture topic, if you wrote it down
- Course level (intro, advanced) — affects how much background to assume
- Your shorthand key (
DDx = differential diagnosis) - A short list of terms you already know, so the model does not over-define them
Copy-ready prompt
Clean these lecture notes for tomorrow's review.
Subject: [area]
Lecture topic: [topic]
Course level: [intro / intermediate / advanced]
My shorthand: [key -> meaning]
Terms I already know: [list]
Notes (verbatim):
"""
[notes]
"""
Return:
1. Hierarchical outline (H2 -> H3 -> bullets) following the lecture flow
2. Key terms in a glossary table: term / one-line definition / where in the outline it appears
3. A "gaps" section: list anything ambiguous in my notes — mark each as [UNCLEAR: ...]. Do not resolve gaps.
4. 5 self-quiz questions covering the most likely exam material, with brief answers
5. A "one-page summary" — only what I would cram on the morning of the exam
Do not invent content beyond the notes. If a step or fact is missing, list it under gaps.A second pass once the outline looks right: “Now generate 10 application questions — not recall — that test whether I can use these ideas, not just remember them.” Recall questions check storage; application questions are what actually transfer to an exam.
If your notes are a photo or a recording
Text in, text out is the cleanest path. Two common starting points:
- Recording. Transcribe first, then feed the transcript into the prompt above. OpenAI’s open-source Whisper is free if you are comfortable with the command line, or its API runs about $0.006/minute (~$0.36/hour) as of June 2026. Otter.ai gives 300 free minutes/month and Notta 120 free minutes/month for live or uploaded audio.
- Handwritten photo. Upload the photo directly — current vision models read handwriting. In 2026 OCR benchmarks, top models land around 90-95% on reasonably clean handwriting, dropping toward ~70% on genuinely messy script. Treat OCR’d text exactly like raw notes: anything the model could not read confidently belongs in your gap list, not silently corrected.
Recommended output structure
H2 headings matching the lecture’s sections, bullets under each, a glossary table, a numbered gaps list with [UNCLEAR] markers, a quiz, and a one-page summary. Avoid prose paragraphs — review notes are scannable, not literary.
How to check the output is usable
- Every key term in the outline also appears in the glossary
- Gaps are flagged, not silently filled (a 90-minute lecture should produce 2-5 gap markers)
- Your shorthand is expanded once, then re-used in the outline
- The self-quiz uses your wording from the notes, not generic textbook phrasing
- Reading the one-page summary tomorrow takes under 3 minutes
If you see zero [UNCLEAR] markers from genuinely messy notes, that is a red flag — the model almost certainly invented its way past your ambiguity. Paste the notes again and add: “Be stricter about gaps; flag anything you had to guess at.”
Common mistakes
- Letting AI fabricate when notes are ambiguous — the most common cause of “I studied this, but the exam asked something different.”
- Skipping the gap list — gaps are your study plan, not a flaw in the output.
- Over-cleaning into a polished essay — unreviewable; you want bullets.
- Mixing two lectures in one cleanup — keep them separate, then merge into a topic map later.
- Asking AI to grade your understanding — it cannot reliably, especially on niche course material.
FAQ
Will the model make up content I didn’t write?
It can, and that is the whole risk here. The prompt counters it twice: it tells the model to flag ambiguity as [UNCLEAR] and forbids resolving gaps. Then you verify by counting gap markers — a real 90-minute lecture should surface 2-5. Claude has a reputation for lower hallucination on messy input, but the instruction matters more than the model.
Which AI is best for cleaning up notes? For one lecture of pasted text, the free tier of ChatGPT, Claude, or Gemini all work. For consolidating a whole course at once, pick a 1M-token model — Claude Pro ($20/mo), or Gemini via Google AI Pro ($19.99/mo). For the lowest chance of fabrication on messy notes, Claude is the safe default (as of June 2026).
What if I only have a recording, not text? Transcribe first, then feed the transcript. Whisper is free (CLI) or ~$0.006/min via API; Otter.ai (300 free min/mo) and Notta (120 free min/mo) are no-setup options.
Will AI handle equations or chemical structures? Inline LaTeX, yes — ask for it explicitly. Structural drawings and complex diagrams, no. Photograph those and keep them alongside the cleaned notes.
Should AI replace my note-taking during class? No. The act of taking notes is part of how memory forms. AI helps after the lecture, not during it.
Related
- Lecture notes cleanup prompts: prompt variants for different lecture formats
- Study plan prompts: turn cleaned notes into a study schedule
- Lecture note cleanup workflow: variant with photo capture and OCR
- Exam study plan: once notes are clean, plan the revision
- AI meeting notes: same approach for meetings, not lectures
Tags: #Workflow #Productivity #Study