How to Prompt AI to Write Speaker Notes

Most AI-generated speaker notes are unusable for one reason: they are paragraphs. You cannot read a paragraph and look at an audience, and a paragraph gives you no idea whether the slide fits the time you have. If you want to prompt AI to write speaker notes you can actually deliver, budget the notes in seconds rather than words, and measure the conversion rate yourself rather than trusting a generic speaking-speed figure.

Measure your own rate first, it takes three minutes

Published speaking-rate averages vary widely and none of them is you. Do this instead.

1. Pick the slide in your deck you understand best.

2. Talk through it once, out loud, as you would on the day. Record the time on your phone.

3. Write down what you actually said, roughly. Count the words.

4. Divide words by minutes. That is your rate.

Most people land somewhere between 110 and 160 words a minute when presenting rather than reading, and the spread within that is enough to ruin a timed talk. Nervous first-slide delivery runs faster than that, which is worth knowing before you write a dense opening.

Now you have a number. A 20-minute talk with 14 slides is 1,200 seconds. Subtract 120 for the opening and closing you will write yourself and 180 for questions and slack. That leaves 900 seconds across 12 slides, and they are not equal: the three slides carrying the argument get 120 seconds each, the rest get about 60.

What to send when you prompt AI to write speaker notes

Give the model the slide content, the per-slide second budget, your measured rate, and the delivery constraints. The budget is the instruction that changes the output most.

Here are my slides as text.

<slides>
Slide 4: "Why reconciliation breaks at scale"
- 3 bullets: volume growth, partial matches, timezone cutoffs
- Chart: match rate falling from 94% to 71% over 18 months

Slide 5: "What we changed"
- 2 bullets: tolerance windows, deterministic ordering
- Code snippet: the matching key
</slides>

Write speaker notes with these constraints:

- Slide 4: 120 seconds. Slide 5: 60 seconds.
- I speak at about 135 words per minute, so budget roughly
  270 words for slide 4 and 135 for slide 5. Stay within 10%.
- Format as short cue lines, max 12 words each, not prose.
  I will be glancing at these, not reading them.
- Each slide's notes must open with the single sentence I say
  while the slide appears.
- Mark one line per slide as CUT FIRST, the line to drop if
  I am running over.
- Do not describe what is on the slide. The audience can see it.
  Tell them what it means.
- No rhetorical questions to the audience.

Two constraints there are doing unobvious work. Max 12 words per line forces cues rather than a script, and a script is the thing that makes presenters stare at their laptop. Do not describe what is on the slide kills the single most common failure of AI speaker notes, which is notes that read the bullets back to you, adding nothing and consuming your whole time budget.

The CUT FIRST marking is cheap insurance. Every talk runs over somewhere, and deciding on stage which sentence to drop is a worse use of attention than deciding it now.

The recovery line

Here is the addition that makes these notes better than anything a generic prompt produces. Ask for the interruption.

For each slide, add one more line labelled IF ASKED.

It should be the question this specific slide is most likely to
provoke from a skeptical person in the room, plus a
20-second answer in cue form.

Base the question on what the slide claims, not on what would
be convenient for me.

On the slide above, that question is some version of how do you know the tolerance window is not just hiding real mismatches, and if you have not thought about it before you are standing up, you will answer it badly. The clause about not choosing a convenient question matters, because a model left to itself will generate the question you would most enjoy answering.

This is the same adversarial move as prompting AI to play devil's advocate, applied one slide at a time where it is specific enough to be useful.

Calibrating the register

Notes written for the wrong audience are the second big failure. The model does not know whether the room is engineers, a board, or a mixed conference track, and it defaults to a register that suits nobody.

State the audience in terms of what they already know rather than their job title. Not notes for executives but they know what a payment reconciliation is, they do not know what idempotency means, do not define either. That instruction is worth more than three sentences about tone, and it is the practical version of what we covered in prompting AI to change reading level.

For a slide where the whole job is making one technical idea land with a non-technical room, generate that slide's notes on its own with the full treatment in prompting AI to explain a technical concept.

The two slides that need different treatment

Generic per-slide notes handle the body of a deck fine and fail on two slides every time.

The question slide, or the end of the talk where questions start, is not a slide at all and does not want cue lines. What it wants is a short list of the questions you expect and the one you are dreading, with the dreaded one answered first in your preparation. Generate it separately, from the whole deck rather than one slide, and ask explicitly for the question you would least like to be asked. Then write that answer out in full, because it is the only place in a talk where a sentence you have already composed beats improvising.

The handover slide, if you are co-presenting, is the other one. Notes written per slide have no idea a different person takes over at slide 8, and the join is where co-presented talks visibly stumble. Tell the model where the handover happens and ask for two lines: the sentence you say to hand over, which should set up what the next person is about to argue rather than just naming them, and the sentence they say to take it. Agree both in advance. Ten seconds of rehearsing that join does more for how a talk reads than another pass on the content.

The pass to run before you rehearse

Notes that survive contact with a stage have had one more thing done to them.

• Read the whole set aloud with a timer, not in your head. Silent reading runs about a third faster than speaking and will tell you the talk fits when it does not.

• Check the transitions. Ask the model for the one sentence that gets you from slide 4 to slide 5, as a separate request. Decks fall apart in the gaps, and notes generated per slide have no reason to join up.

• Delete the first line of every slide if it names the slide. It will sometimes do this anyway despite the instruction, and the audience has already read the title.

• Keep the CUT FIRST lines visible rather than deleting them during edits. They are only useful on the day.

If the deck is an investor update rather than a talk, the notes and the document want different structures entirely: prompting AI to write an investor update covers that case. The underlying techniques, examples and explicit constraints, are the ones in our prompt engineering pillar, and giving the model a sample of notes you have delivered well is the fastest way to fix a register you cannot describe.

Frequently asked questions

How many words per minute should I budget?

Measure rather than assume: time yourself on one slide and divide. Presenting rates commonly fall between 110 and 160 words a minute, which is a wide enough band that using the midpoint for a 20-minute talk can leave you three minutes out.

Should speaker notes be full sentences?

Only the first line of each slide, the one you say as the slide lands. Everything after that should be cue lines of a dozen words or fewer, because full sentences invite reading and reading loses the room.

Can AI write the slides as well as the notes?

It can, and the notes come out better when it does not. Notes generated from slides you wrote are anchored to an argument you already decided on. Notes generated alongside freshly invented slides tend to be confident about claims you cannot support.

What about notes for a pre-recorded talk?

Then you want a script, not cues, and you should say so. Drop the 12-word limit, keep the second budget, and drop the IF ASKED lines since nobody can interrupt a recording.

How do I stop the notes just repeating the bullets?

Instruct it explicitly not to describe the slide, and then check. This instruction gets ignored more than any other in the prompt, and the fix is to delete the offending lines and regenerate that slide alone rather than rephrasing the rule.