Which tool to choose
They're not equivalent. The choice depends on how much material you have and how much you fear errors.
Short notes, already in text format (phone notes, scattered lines, email drafts) → ChatGPT or Claude, free version. You paste, ask for the structure, copy the result. It's the fastest route.
Many separate documents (PDFs, slides, meeting minutes, recordings) → Google NotebookLM, free with a personal Google account. In the free version you can upload PDFs, web pages, Google documents and presentations and YouTube videos, and generate summaries, FAQs, timelines and briefing documents. The advantage that counts: NotebookLM cites the exact sentences from your sources, so you verify every statement with a click.
Notes you still have to "get out" by speaking → the ChatGPT app's voice dictation (microphone icon), then the rewrite in the same chat. You speak, it transcribes, you ask it to structure.
Between ChatGPT and Claude, for a long text Claude produces more orderly prose, and its free version has a detail that counts: it includes web search and file upload (up to 20 files per chat and 500 MB per file), plus Projects and Artifacts. Free ChatGPT is tighter on files: free users get about 3 files a day.
How to do it
The fastest route starts from ChatGPT or Claude. From browser or app, the steps don't change.
- Open a new chat: a clean chat prevents old instructions from contaminating the result.
- Gather all the notes into a single block. Even if they're chaotic, fragmentary, full of abbreviations: they're fine like that.
- Paste the notes after the prompt you find below, in the same message.
- Send and read the draft. Almost always a second round of refinement is needed (see point 6).
The syntax to paste before your notes:
Turn the notes below into a structured and readable document.
Rules:
- Use exclusively the information present in the notes, don't add anything of your own.
- Create a title, subheadings for each theme and bullet points where useful.
- Group similar ideas even if in the notes they're far apart.
- Keep a professional and flowing tone.
- If you find an unclear or contradictory point, flag it at the end in a "To verify" section.
Notes:
[paste everything here]
- Check: the final "To verify" section is your quality control. If it's empty, the AI understood everything; if it contains points, those are the passages where you risk a fabrication.
- To refine, stay in the same chat and ask for targeted changes, one at a time: "Shorten the introduction", "Turn section 2 into a table", "Make the tone more formal".
If instead you upload documents to NotebookLM, create a new notebook and add the sources (PDFs, web pages, YouTube videos or pasted text). Then, in the notebook chat, give the same structure request. Avoid generic questions like "explain everything": ask something precise, for example "explain the causes using only the uploaded PDF".
Concrete example
Marco, a real estate agent, comes out of a viewing with these notes dumped on his phone: "apartment via Dante 3rd floor no elevator — 85 sqm — kitchen redone 2022 — bathroom to fix — dual exposure — independent heating — building fees €90/month — courtyard view — owner asks 215k negotiable — available from September — energy class D".
He opens ChatGPT, pastes the prompt above and adds the notes. In a few seconds he receives a structured listing: a title ("Bright 3-room apartment of 85 sqm on via Dante"), a "Features" section with a bullet list (size, exposure, independent heating, building fees), a "To assess" section with the bathroom to renovate and the absence of an elevator, and at the bottom the "To verify" section with the note: "Confirm whether the negotiable price of €215,000 should be shown in the listing or only internally". Marco corrects one line, removes the price from the public listing and in two minutes he has a text ready to publish, started from fifteen disconnected words.
When it does NOT work (and how to fix it)
If the AI makes up details you didn't write
It's the number-one risk: it fills the gaps in the notes with plausible but false information. The prompt above limits it with the rule "use exclusively the information present". For a stronger guarantee, move the work to NotebookLM: it uses a system (called RAG) that anchors the answers to your sources and adds citations to the exact passages, so fabrication has less room. Even NotebookLM, though, can make mistakes or misinterpret a passage. Always check numbers and proper names against the original.
If you have too many files and free ChatGPT blocks you
On the free plan the wall comes fast: 3 files a day, with a reset every 24 hours. Two ways out without paying: copy the text of the documents and paste it directly into the chat (the limit applies to uploaded files, not pasted text), or switch to NotebookLM, much more generous on sources. NotebookLM's free version allows 50 sources per notebook and up to 100 notebooks.
If the notes are an audio recording
ChatGPT alone digests audio poorly: it transcribes with the Whisper model, but with a 25 MB limit per file, no speaker distinction and no integration with meetings. The clean route is NotebookLM, which transcribes audio and video files automatically: you upload the recording of a speech or an interview and it gives you back the text, on which you then ask for the structure.
If the text comes out flat and all the same
Almost always the problem isn't the AI, it's the too-vague prompt. Specify recipient and purpose: "This text is for a client who doesn't know the subject" produces a different result from "This is an internal reminder for my team". On NotebookLM you can also change the instructions from the notebook settings, where you choose the tone (formal, technical or simpler) and the level of detail (from a bulleted summary to a full deep dive).
A tip from someone who actually uses it
Don't aim for the perfect text on the first shot: aim for a rough but complete draft, then work it piece by piece. The most efficient flow is hybrid: voice to capture, text to refine. An entrepreneur records ten minutes of voice notes while driving; the AI transcribes, summarizes and divides the ideas into categories; in front of the PC he refines and turns it all into an official document. Dictating is much faster than writing, and this changes the amount of material you manage to gather before structuring it.
Frequently asked questions
Can I do all this for free?
Yes. Claude has a free plan that includes the Sonnet model, file upload and basic writing functions, with a cap of about 20-30 messages during peak hours. NotebookLM is free with a Google account. Free ChatGPT works great for pasting text. You pay only if you work at high volumes every day.
Do my notes stay private?
On NotebookLM yes by default: Google doesn't use them to train its models, the information you upload stays yours. On free ChatGPT and Claude, instead, the conversations can end up in training: open the privacy settings and disable sharing if the notes are confidential.
How much material can I paste at once?
More than you think. Claude's free version has a 200K-token context window, enough for long documents, entire codebases or hefty research articles. In practice you paste dozens of pages of notes in a single message.
Do I have to give ChatGPT a complicated prompt or is it enough to ask "fix these notes"?
"Fix these notes" works too, but the result is a lottery: sometimes orderly, sometimes invented, sometimes off-tone. The difference isn't in the complexity of the prompt, it's in two precise instructions: "use only the information present" (blocks fabrications) and "the text is for [recipient]" (decides the tone). Two extra lines save you three rounds of corrections.
Isn't it safer to write the text myself instead of trusting the AI?
False dilemma. The AI doesn't replace your judgment, it gives you a starting point in thirty seconds instead of half an hour. The control stays yours: you reread, correct the facts, decide what to keep. The real risk isn't that the AI writes badly, it's that you publish without rereading. The study rule applies: for a review it's perfectly fine, for important things you always go back to the original source.