How to Write a Speech With AI

The cursor blinks. In three weeks you are standing up in front of a room full of people who all know the person you are about to talk about, and right now you have nothing. A wedding toast, a leaving do, a graduation, a 70th birthday. The stakes feel enormous and the page is empty.

So you do the obvious thing. You open an AI, type “write me a best man speech,” and thirty seconds later you have three warm paragraphs about the happy couple, the journey of love, and how you always knew they were meant to be. It reads fine. It also reads like a greeting card anyone could have written, and the room will feel that the moment you start talking.

There is a better way to use AI here, and it is not “write my speech.” It is closer to “help me shape the speech I already have in my head.” You bring the real memories and the real feeling. The AI handles the part that makes people freeze: turning a pile of half-formed thoughts into something with a clear beginning, middle, and end. Here is how to do it, move by move.

Dump your real memories before you touch a prompt

The single biggest reason AI speeches fall flat is that people ask for the speech before they have given the AI anything true to work with. With nothing to go on, the model fills the gap with sentiment, because sentiment is free and specifics are not.

So do the specifics first, before any clever prompt. Open a voice note and just talk for two minutes, as if a friend asked “so how do you know them?” Say the day you met, the funny thing they did on holiday, the moment you realised how much they meant to you, the phrase they always say. Then paste that rambling transcript into the AI. Messy is fine. Messy is real, and real is the whole point.

Give it the occasion, the room, and the clock

Once it has your raw material, your prompt is not “write a speech.” It is a short brief. Tell the AI who is speaking, who is listening, what feeling you want in the room, and how long you have. That last one matters more than people think. Most wedding planners put a toast at two to five minutes, and since we speak far slower than we read, roughly 120 to 150 words a minute, that is only about 300 to 600 words. A speech that overruns is the one people remember for the wrong reason.

Ask for spoken rhythm, not written prose

A speech is not an essay you happen to read out loud. The sentences have to work when spoken, by a slightly nervous person, to a room that only hears them once. AI defaults to written English: long clauses, tidy lists, commas everywhere. Left alone it gives you something that looks good on the page and falls flat when you say it out loud.

So ask for the spoken version explicitly. Tell it: short sentences, one idea each, contractions, a couple of deliberate pauses, nothing you would trip over. Then read it back out loud. If a line makes you run out of breath or sound stiff and formal, it is written, not spoken, and it needs cutting down. If the whole thing still feels too formal, our guide on how to make AI writing sound human covers the exact tells to strip out.

Build it around one real story, not three

The instinct is to cram in everything: the school days, the road trip, the time with the canoe. Three half-told stories always lose to one told properly. A speech that lands has a spine, usually a single scene the audience can picture, and everything else hangs off it.

Pick the one memory that says the most about who this person is, and ask the AI to give it room. The shape underneath almost every good short speech is the same four beats, so hand the model that structure along with your material.

Give it those four boxes and your notes, and it will stop wandering and start building something that actually goes somewhere.

Cut the greeting-card lines

Even with good material, models drift toward the phrases that make a speech feel generic. Watch for “little did I know,” “words cannot describe,” “on this special day,” “the journey we’ve been on,” and any sentence that would fit literally any couple or any retiree. A line that fits everyone fits no one.

The fix is fast: ask the AI to reread its own draft and flag every line that could have been written about a stranger. Then you decide what stays. What is left should be things only true of this person. If you want a final tidy-up for the small stuff, a quick pass through a grammar check catches the typos before you print your cue cards.

Read it aloud and time it, twice

This is the step everyone skips and the one that turns an AI draft into your speech. Stand up. Read it out loud, properly, as if the room is there. Time it on your phone. You will feel three things instantly: the lines that are too long, the jokes that need a beat before the punchline, and the bits that are not you.

Change those bits. Swap the AI’s word for the word you would actually use. Add the aside you always make. This back-and-forth, where you read and the AI reshapes, is worth doing two or three rounds. If you want a guided version of that loop instead of a blank chat, Chatday’s help me write tool walks you through the brief and the redrafts, so you supply the memories and it handles the structure.

Which AI should you use for it?

Any of the strong general models write a decent speech, and each has a slightly different ear, so it is genuinely worth running your notes through two and keeping the version that sounds most like you. Claude Opus 4.8 is good at a warm, natural tone that does not get sentimental. Gemini 3.1 Pro is strong at pulling the one usable line out of a long, rambling voice note. Having them side by side means you paste the same memories into each without hopping between apps and logins, then keep the best line from each.

So before you touch a single sentence, do the one thing the whole method rests on. Open your phone tonight, hit record, and talk for two minutes about the person, the way you would tell a friend. That rough voice note is most of the speech already. Everything after it is just shaping, and that is the part you can hand over.