Quick answer
Here is how to write speaker notes using ChatGPT that sound spoken rather than written: ask for the exact words you will say, give each slide three beats — an opening line, the meaning the slide cannot show, and a bridge onward — then cap the words to your slide timing.
Learning how to write speaker notes using ChatGPT takes about fifteen minutes. Learning how to get notes you can actually say out loud, standing in front of your director, without sounding like you are reading a press release — that takes one specific change to the prompt.
Here is what goes wrong when you skip that change. You paste your slide bullets in, ask for speaker notes, and get back something that is technically correct and completely unsayable.
- Sentences with three clauses that leave you out of breath halfway through
- Notes that just read the slide back to you, word for word
- Two hundred and eighty words for a slide you have forty seconds to cover
The fix is not a better prompt template. It is a structural one: your slides are written for the eye, your notes have to be written for the ear, and ChatGPT will default to the first one every single time unless you stop it. This guide gives you the system that stops it — the Three-Beat Note, the approach we teach at PromptPeakAI for turning slide bullets into a spoken script.
Will this work for you?
Why ChatGPT Speaker Notes Sound Written
ChatGPT sounds stiff in speaker notes because it is producing written English by default, and written English is a different language from spoken English. Written prose can carry subordinate clauses, semicolons and formal connectives, because your reader controls the pace. A listener does not. They get one pass, at your speed, with no way to re-read.
Most professionals never name this problem, so they blame the tool and go back to writing notes by hand. The gap is real, and it is visible the moment you put the two registers side by side.
✕ Written register — what you get by default
"This quarter's performance demonstrates a sustained upward trajectory across all three regions, with EMEA in particular exhibiting growth that exceeded our initial forecasting assumptions by a considerable margin."
✓ Spoken register — what you can actually say
"All three regions grew this quarter. EMEA is the one to watch. We forecast eight percent there. It came in at fourteen. I want to spend a minute on why, because it changes what we do next quarter."
Read both aloud. The first one takes a full breath and lands as a paragraph. The second one lands as four thoughts, and the audience keeps up with all of them. Same facts, different job.
The Three-Beat Note Explained
The Three-Beat Note is a structure for speaker notes in which every slide’s notes contain exactly three beats: an Open that earns attention, a Carry that delivers the meaning the slide itself cannot show, and a Hand-Off that walks the room into the next slide. Three beats, in that order, every time.
The Three-Beat Note: what each beat does, and roughly how much of your talking time it should take.
The Carry beat is where the whole thing lives or dies. Your slide already shows the number. Your voice has to explain why the number moved, what it means for the people in the room, or what you want them to do about it. If your notes and your slide say the same thing, you have made yourself optional.
This is the Logic System’s Role, Goal, Context and Format pointed at one narrow job. The Three-Beat Note is what the Format element looks like when the output is meant to be spoken rather than read.
How to Write Speaker Notes Using ChatGPT
The workflow is four steps and it runs in a single chat. Prime the deck once, write one slide’s notes, tighten to a word budget, then say it out loud before you accept it. After the first slide, the rest of the deck goes fast.
The four-step loop. Step one happens once per deck; steps two to four repeat per slide.
Prime the deck once
Before you ask for a single note, tell ChatGPT what the whole talk is for. Skipping this is the most common reason notes drift off-message by slide six — the model has no idea where the deck is going, so it treats each slide as a standalone essay. Good context up front costs you one message and saves you the whole deck.
You are a speech ghostwriter helping me prepare to present. Do not write anything yet — just hold this context for the rest of our conversation. The deck: [what it is, e.g. Q3 operations review] The audience: [who is in the room and what they care about] The room: [e.g. 12 people, video call, cameras on] Total speaking time: [e.g. 18 minutes across 14 slides] What I want them to do afterwards: [the one decision or action] How I talk: direct, plain words, short sentences, no buzzwords. Confirm you have this, then wait for my first slide.
Write the three beats
Now paste one slide’s text and ask for the beats by name. Naming the structure is what stops the model reverting to essay mode — you are specifying the output format rather than hoping for it.
Here is the text on slide [number]: [PASTE SLIDE TEXT] Write my speaker notes for this slide as three beats, labelled: OPEN — one sentence that earns their attention. Never a summary of the slide title. CARRY — two to four sentences giving the meaning the slide cannot show: why it moved, what it means for this audience, what is at stake. HAND-OFF — one sentence that leads into the next slide. Not "moving on to". Rules: - Write the exact words I will say out loud, not a description of what to cover. - Short sentences. Contractions. No sentence longer than 20 words. - Do NOT repeat anything already printed on the slide. - Do NOT invent numbers, dates, names or results. If you need a fact I have not given you, write [CHECK] instead.
Set the word budget
Unbounded, ChatGPT will hand you two hundred and eighty words for a slide you meant to spend forty seconds on. Give it a number instead of a vibe. The pacing table in the next section tells you which number.
Read it aloud before you accept it
This step is not optional and it takes ninety seconds. Read the notes out at presenting speed. Anywhere you stumble, run out of breath, or hear yourself sounding like a brochure, that sentence is written not spoken — send it back. Fix only the beat that failed rather than regenerating the whole thing.
You should now have notes for one slide that you can read aloud in a single comfortable pass, with nothing in them that also appears on the slide.
Paste the result into the notes box under the slide. In PowerPoint that is the Notes pane below the slide area, and it shows up beside your slides in Presenter view while the audience sees only the deck — Microsoft’s guide to the Notes pane covers the view settings. Google Slides works the same way, in the panel underneath the canvas.
The PowerPoint Notes pane below a slide, with three-beat speaker notes pasted in and the Notes button visible on the status bar.
Crop to the slide area and Notes pane only. Hide the file path, account name, and any real client data on the slide.
File: powerpoint-notes-pane-paste.png · Alt: "Three-beat speaker notes pasted into the PowerPoint Notes pane below a slide"Match Word Count to Speaking Time
A comfortable presenting pace is around 130 words a minute — slower than conversation, because you are pausing for emphasis and letting things land. That number turns your slide timing into a hard cap you can put in the prompt.
| Time on the slide | Words to aim for | Line to add |
|---|---|---|
| 30 seconds — a single stat | 65 to 75 | Limit these notes to 75 words maximum |
| 60 seconds — a standard slide | 130 to 140 | Limit these notes to 135 words maximum |
| 90 seconds — a chart or model | 195 to 210 | Limit these notes to 200 words maximum |
| 2 minutes — a story or case | 260 to 280 | Limit these notes to 270 words maximum |
Spoken word budget by slide time, calculated at a presenting pace of roughly 130 words per minute.
Add the cap to the end of the master prompt and the whole tone shifts. Constrained, ChatGPT drops the qualifying clauses first — which are exactly the parts that made the notes sound written. Most guides on how to write speaker notes using ChatGPT skip this entirely, and it is the single cheapest improvement available to you.
Prompts for Three Common Slide Types
Three slide types break the standard prompt, and each needs one extra instruction. Chart slides, transition slides, and any slide where sounding like yourself actually matters.
Chart and data slides
The failure here is reading the axes out loud. Your audience can see the axes. What they cannot see is which direction is good.
This slide is a [chart type] showing [what it plots]. The key numbers are: [list them]. Write the CARRY beat only. Do not describe the chart. Instead: 1. Name the single thing I want them to notice. 2. Explain in plain words why it happened. 3. Say what it means for [audience] in the next quarter. Then add one bracketed stage direction telling me where to pause and point. Limit to [number] words.
Transitions between slides
Most decks die in the gaps. “Moving on to the next slide” is not a transition, it is an admission that there is no argument connecting the two.
Here is slide [N]: [PASTE] Here is slide [N+1]: [PASTE] Write only the HAND-OFF beat for slide [N] — one sentence that makes slide [N+1] feel like the answer to a question slide [N] just raised. Banned openings: "moving on", "next up", "now let's look at", "as you can see".
Sounding like yourself
If people know how you talk, generic AI cadence is obvious. The fix is to show ChatGPT a sample rather than describe your style — the same few-shot approach that works for written voice matching.
Below is a transcript of me talking for two minutes about a work topic. It is unpolished on purpose. [PASTE YOUR OWN TRANSCRIPT — dictate it into your phone's voice notes and paste the text] Study my sentence length, my filler habits, the words I actually use, and how I open a point. Then rewrite the speaker notes below in that voice. Keep my quirks. Do not upgrade my vocabulary. [PASTE THE NOTES TO REWRITE]
Paste Slide Text or Upload Decks
Both routes work, and the right one depends on how big the deck is and how much of it is pictures. Pasting text gives you control slide by slide; uploading the file is faster for long decks but comes with a limitation almost nobody warns you about.
Choosing between pasting slide text and uploading the deck file.
Here is the limitation. When you upload a presentation or a PDF, ChatGPT extracts the digital text and discards the images — OpenAI’s own file-upload documentation states that outside Enterprise, document uploads are text-based retrieval only. Your revenue chart, your process diagram, your org structure: invisible. The notes come back confident and silent about the one slide you most needed help with.
I have uploaded my slide deck. Two things before we start: 1. List the slide numbers where the text you extracted looks thin or missing — those are probably images, and I will describe them to you. 2. Do not write notes for any slide until I confirm what is on it. Then wait. I will give you slides one at a time.
The ChatGPT message composer with a presentation file attached, showing the attachment chip above the input box.
Crop out the conversation sidebar, account name, and browser bookmarks. Blur the deck filename if it names a client.
File: chatgpt-upload-presentation-deck.png · Alt: "A presentation deck attached to the ChatGPT message composer before generating speaker notes"On plan choice: pasting text works on any account, including free. Uploading is where the tiers bite — free accounts are capped at three file uploads a day, which is one deck and a mistake. If deck uploads become part of your week, that is the argument for paying, and our plan comparison covers the rest of the trade. Teams already inside Microsoft 365 should also weigh Copilot working directly inside PowerPoint, which skips the copy-paste entirely.
Before you upload a client deck
A slide deck is one of the most sensitive things you will ever paste into an AI tool. It has unreleased numbers, named accounts, and sometimes salary bands. The account type matters more than the price here: consumer plans use your content for training by default with an opt-out, while business and enterprise accounts do not.
Public messaging, your own outline, anonymised figures.
Client names, internal targets, anything naming a colleague.
Unreleased financials, personal data, pre-announcement material.
If you are unsure where your account sits, check the data controls before the deck goes anywhere near the upload button — our workplace privacy guide walks through the settings.
When AI Speaker Notes Go Wrong
Four failures account for nearly everything that goes wrong here, and each has a one-line fix you add to the prompt rather than a whole new approach.
| What you see | Why it happens | The fix |
|---|---|---|
| Notes read the slide back | No rule separating slide from voice | Add: do not repeat anything printed on the slide |
| You run out of breath | Written syntax with stacked clauses | Add: no sentence longer than 20 words |
| Every slide runs long | No word cap in the prompt | Add the word budget for that slide's timing |
| It invents a result | Thin context, model fills the gap | Add: write [CHECK] instead of guessing |
That last row is the one that can genuinely embarrass you. A model asked to sound confident about a slide it only half understands will produce a plausible number, and a plausible number in front of a finance director is worse than an honest gap. This is ordinary AI hallucination wearing a suit — every figure in your notes gets checked against the source before you present.
✕ Notes that echo the slide
"As you can see on this slide, Q3 revenue was 4.2 million, which represents an increase over Q2, and our three key drivers were pricing, retention, and the new mid-market segment."
✓ Notes that carry the meaning
"Four point two million. The number is on the slide, so I won't read it to you. What matters is where it came from — mid-market, which we only opened in May. That changes the case for headcount, which is the next slide."
What this will not do for you
Being straight about the limits is what keeps you out of trouble on stage. ChatGPT is writing a script from what you gave it, and there are things it structurally cannot know.
Four things stay yours, and outsourcing them is how presentations go badly:
- The room's history. AI does not know that this proposal already failed once in March, or which VP takes the mid-market question personally.
- Every number you say out loud. Verify each figure against your own source. The notes are a draft, not a fact-checked script.
- What you actually believe. If your recommendation is genuinely uncertain, the notes should say so. AI defaults to confident.
- The delivery. Nothing here fixes pace, eye contact or nerves. It gives you words worth saying, which is a different problem from saying them well.
Do one slide with this system before you commit to the whole deck — the middle slide, not the title slide, so you are testing it on something real. If the notes for that slide read aloud cleanly on the second attempt, the rest of the deck is twenty minutes of work rather than an evening of it. That is really all there is to how to write speaker notes using ChatGPT: prime once, three beats, a word cap, and your own voice reading it back before anyone else hears it.
Write one slide's notes before your next deck
- Prime ChatGPT once with your audience, room, total time and the outcome you want. Paste one slide and ask for three labelled beats: Open, Carry, Hand-Off. Add a word cap that matches your slide timing — roughly 135 words per minute on screen. Ban it from repeating anything already printed on the slide. Read the notes aloud and fix only the beat you stumbled over. Check every number in the notes against your own source before you present.