You have a speech to give next Tuesday. The draft is done — it is a real story about your life, something that actually happened. You open your favorite AI tool, paste it in, and ask for "feedback." Ten seconds later you get back a version that is grammatically flawless, better organized, and reads like a press release written by a committee. The problem: it is no longer your speech. The jokes are gone. The stumble where you admit you were scared is gone. The line your daughter says when she is six years old is now "the child expressed enthusiasm." Nobody talks like that on stage, and nobody wants to listen to it.
This is the core tension of using AI for public speaking. The tool is genuinely useful — it can catch weak transitions, flag sections that run long, and point out where a story has no payoff. But if you let it "improve" the prose, you end up with a polished speech that sounds like everyone else's. The fix is a mindset shift: treat AI as your rehearsal coach and evaluator, not your ghostwriter. Give it a rubric and a stopwatch, not a blank page and a thesaurus.
I give speeches in my Toastmasters club in Kluang, Malaysia, and I have landed on a workflow that keeps AI in the helper seat. Here is how it works.
Start with the right raw material
The single most important rule is this: the AI must work from your words, not generate them for you. If you paste in a one-line idea and ask the AI to "write a five-minute speech," you have already lost. The stories, the voice, and the emotional arc have to come from you first. AI cannot remember the smell of your grandmother's kitchen or the panic you felt the first time you stood in front of a client meeting. It can only rearrange what you give it — and if you give it nothing, it fills the gap with generic filler.
So begin with a real draft. It does not have to be good. In fact, a messy first draft is ideal, because it is full of your natural phrasing. Write it the way you would say it out loud, complete with the words you would actually use in a conversation with a friend. Do not edit as you go. Get the story down in one sitting, even if it is rough.
Then give the AI context, not just text. A good opening prompt looks something like this:
I'm a Toastmasters member preparing a 5-to-7 minute speech. I've attached my rough draft. I want you to act as my speech coach, not my ghostwriter. Please do NOT rewrite my speech or improve my phrasing. Instead, evaluate it against a clear set of criteria and give me specific, actionable feedback.
That last sentence matters more than you think. Most AI tools default to helpful over-rewriting, because that is what most users ask for. You have to explicitly take the red pen away from it and hand it a clipboard instead.
Give it a real evaluation rubric
Toastmasters has a long tradition of evaluating speeches, and the evaluation is structured. There is a judging criteria sheet used in contests, and the core categories are: speech value, organization, and — for the delivery side — voice, gestures, and language. When you hand the AI a concrete rubric, its feedback becomes dramatically more useful than a generic "great speech!" or "consider shortening your introduction."
Here is a rubric prompt you can copy and adapt:
Evaluate my draft against these Toastmasters-style criteria, and score each on a 1–10 scale with one concrete suggestion per point:
- Opening — does it hook the audience in the first 20 seconds?
- Organization — is there a clear through-line from opening to conclusion?
- Speech value — does the audience leave with a clear takeaway?
- Stories — does each anecdote have a setup and a payoff?
- Transitions — are the jumps between sections smooth?
- Word economy — are there sentences that can be cut without losing meaning?
- Vocal variety potential — where are the moments that could use a pause, a change of pace, or emphasis?
- Conclusion — does it land, or does it fade out?
When you score it this way, something useful happens: the AI stops trying to be your editor and starts behaving like an evaluator. It points at specific paragraphs and says "this transition is abrupt" or "the payoff in story two comes too late." That is exactly the kind of note a human evaluator would give you on a slip of paper after your speech — except you get it before the meeting instead of after.
Three passes, each with one job
Once the raw evaluation is done, I run the draft through three focused passes. Each pass has exactly one job, and I never ask the AI to do two things at once. This keeps the output controllable and stops it from drifting into a full rewrite.
Pass one — structure and clarity. Ask the AI to identify the weakest transition and the least clear paragraph, and to suggest one way to fix each — but in your words. A good instruction is: "Suggest a fix, but keep my sentence structure and vocabulary. Do not replace my opening line."
Pass two — timing and length. A 5-to-7 minute speech is roughly 650 to 950 words when spoken at a natural pace. Ask the AI to estimate the speaking time of each section, flag any section that runs long, and suggest where to cut. This is where the stopwatch role comes in. Most of my first drafts run about thirty percent too long, and the AI is very good at spotting which paragraph is the one that can go.
Pass three — delivery notes. This is the part people forget. Ask the AI to read the draft as if it were being spoken and mark where a pause would land, where the pace should slow down, and where you should let the audience react. I ask for delivery notes in brackets, like this: [pause], [slow down], [smile]. These are stage directions, not prose changes, so they cannot dilute your voice.
Running these as three separate prompts — rather than one big "make it better" — is the difference between a useful session and a frustrating one. When you give the AI a single job, it does that job well. When you give it three jobs, it averages them into mush.
Protecting your voice: the "flag, don't rewrite" rule
Even with a rubric and a single job, the model will occasionally slip back into improving your sentences. You need a standing rule that survives across all three passes. I call it the "flag, don't rewrite" rule, and I paste it at the top of every prompt:
Flag problems and suggest options. Never rewrite my sentences. Keep my anecdotes, my jokes, and my exact wording for any quoted dialogue. If you must show me a change, show it as an optional suggestion in brackets, never as a replacement.
Quoted dialogue deserves special protection. The line your child actually said is untouchable, because it is true in a way no polished paraphrase can match. If you let the AI "clean up" a real quote, you have invented a detail in your own life story, and the audience can usually tell. Real quotes have a specific, awkward rhythm that sounds like a real human. Smooth quotes sound like marketing copy. Keep the awkward ones.
The same goes for jokes. Humor is fragile. A joke that works depends on exact timing and exact wording. An AI will happily "improve" a punchline into something more logical — and logic is the enemy of funny. If the AI flags your joke as "unclear," read it out loud to a friend before you change a single word. More often than not, the delivery needs work, not the line.
Rehearse with it as a pacing coach
The most underrated use of AI in speech prep is rehearsal feedback on delivery, not content. Here is a simple loop that works well:
- Read your speech out loud and record it — your phone's voice memo app is fine.
- Note your actual time. If you are aiming for 5-to-7 minutes and you land at 8:30, you now have a concrete cutting target.
- Ask the AI: "Here is my speaking time. Which section should I shorten, and by roughly how many words, to get under 7 minutes?" Word counts map to time, so this becomes a quick arithmetic problem the AI handles cleanly.
If you want to go further, you can transcribe a practice run (many AI tools will transcribe audio now) and ask for a filler-word report — how many "um," "so," and "you know" slipped in. One pass of this will cut your filler words by more than any amount of self-editing, because you will finally see the habit instead of just feeling it.
There is also a classic Toastmasters skill called "vocal variety" — the use of pace, pitch, volume, and pauses. Ask the AI to mark your script with where those changes should happen. A script with stage directions is a genuinely different thing from a script without them, and it makes your rehearsal time far more efficient.
What AI should never do for your speech
For all its usefulness, there are lines the tool should not cross, and being clear about them keeps you honest.
It should never write your opening story. The opening is your moment to be real. An AI-generated opening anecdote is a generic one by definition — it is assembled from the most common patterns, which means it is the least likely thing to surprise a room.
It should never invent facts about your life. If the AI "adds a nice detail" to a story, that detail did not happen. On stage, small invented details have a way of collapsing when someone asks a follow-up question. Keep your memory the sole source of truth.
It should never replace your memory of how the speech felt. This is the big one. AI works with text, but a speech is a live performance. It cannot hear the room go quiet during your quiet moment, cannot feel when the audience is with you, and cannot tell you to slow down because you are nervous — only you and your evaluators can do that. Use AI to prepare the map, but you still have to walk the route.
Putting it together
The workflow I land on, end to end, is this: write a messy first draft in your own voice, then run one evaluation against a clear rubric, then three single-purpose passes for structure, timing, and delivery notes, then record yourself and use the transcript for filler words and pacing. At every step, keep the "flag, don't rewrite" rule in place.
What you get out is still your speech — your stories, your jokes, your awkward real quotes — but it is tighter, better timed, and more confident on the page. What you do not get is a soulless block of text that could have been written by anyone. The difference is the difference between a speaker and a robot reading a script, and in a Toastmasters meeting, that difference is everything.
The next time you are staring at a rough draft the night before a speech, try handing the AI a rubric and a stopwatch instead of a blank page. Keep your voice, and let the machine do the bookkeeping. That is the whole trick — and it works.
