How to Use AI to Write a Speech in 2026: Step-by-Step Guide with Prompts

If you are wondering how to use AI to write a speech in 2026, the short answer is a six step workflow: frame the audience and goal with ChatGPT, gather evidence with Perplexity, draft with Claude, edit with Grammarly and Wordtune, rehearse with an AI voice from ElevenLabs, and build the supporting deck with Gamma. Based on testing this stack across speech types from wedding toasts to investor keynotes, the full workflow takes about three focused hours, and every step below includes a copy-paste prompt.

• The whole workflow fits in one afternoon. Briefing in ChatGPT takes 20 minutes, research in Perplexity takes 30, the draft in Claude takes 40 including one revision round, editing in Grammarly and Wordtune takes 40, and the rehearsal pass with ElevenLabs plus a slide deck in Gamma closes out the remaining hour.
• Length is arithmetic, not guesswork. Natural delivery runs at 140 to 150 words per minute, so a 5 minute speech is 700 to 750 words and a 10 minute keynote is 1,400 to 1,500 words; setting the word count in the draft prompt is what keeps the final speech inside its time slot.
• It costs nothing to start. Every tool in this guide offers a genuine free tier, and in the AITokenHub database 206 of the 300 tracked tools offer free plans, so the required investment is your time rather than your budget.
• Top rated picks for this task: ChatGPT leads the stack at 4.7/5 in our database, Claude is the strongest long-form draft writer at 4.6/5, and ElevenLabs matches it at 4.6/5 for the rehearsal step that most speakers skip and most audiences notice.
• AI writes the draft, but you own the delivery. The model supplies structure and flow, while the stories, the facts, and every claim must come from you, because an invented statistic in a live speech is a credibility problem no edit can fix.

Speeches are the one writing format most people practice too rarely to master, and the fear data is blunt about it. The widely cited estimate from the National Institute of Mental Health puts glossophobia, the fear of public speaking, at roughly 73 percent of the population, which makes stage fright one of the most common fears measured in communication research. The anxiety is usually about two failure modes: losing the thread midway, and running long or short against the clock. A structured AI workflow attacks both directly, because the draft gives you the thread and the word count math plus a timed rehearsal gives you the clock.

What changed in 2026 is the quality of the first draft. McKinsey State of AI research reported that 78 percent of organizations now use AI in at least one business function, and writing assistance remains among the most adopted uses because drafting is exactly what modern models handle best. A speech that once meant a blank page and three evenings now means a clear brief, one strong prompt, and a revision pass, which is why this workflow compresses weeks of spare-time drafting into a single afternoon session. The tools are also no longer exotic: the writing and audio categories in our database alone hold dozens of free-tier options, and the six tools in this guide cover every stage from framing to rehearsal.

There is a second advantage that has nothing to do with speed: iteration. A human speechwriter hands you one draft, but asking Claude for three different openings, or for the same point made with a story instead of a statistic, costs one prompt each. That means you can test an opening aloud before choosing it, which used to be a luxury only professional keynoters had. The rest of this guide shows the full workflow step by step, with the exact prompts to copy.

The economics of the old way explain the adoption curve as well. Professional speechwriting runs anywhere from a few hundred dollars for a single wedding toast to several thousand for a corporate keynote, and most people who give two speeches a year could never justify it. The tools in this guide do the structural work of that process for the price of an afternoon, and the free tiers alone cover the occasional speaker completely, which in our database describes 206 of the 300 tracked tools. That combination of zero cost and same day turnaround is why AI drafting has become the default first step for speeches at every level, from school competitions to sales kickoffs, while the craft of deciding what to say stays exactly where it always was, with the speaker.

This step produces the speech brief, a short document that defines who is listening, what the occasion demands, and what the single takeaway should be, and every later step quotes it back, so doing it first is what keeps the draft from sounding generic. First, open ChatGPT and enter this prompt, replacing the bracketed parts with your details:

Act as a speech coach. I am giving a speech and need to define it before drafting. Interview me one question at a time about: who the audience is and what they expect, the occasion and the tone it calls for, the single takeaway I want them to remember a week later, how long I speak, and any stories or facts only I know that must be included. After my answers, write a speech brief of 150 words or fewer that locks in: audience, occasion, tone, takeaway, time limit, and three must-use personal details. Then list three opening angles I could take.

Answer the questions honestly and specifically, because the brief is the highest leverage 20 minutes in the whole workflow. When ChatGPT proposes the three opening angles, pick one and say why in a follow-up message, for example: use the second angle, the story about the failed launch, because this audience of founders will recognize themselves in it. That kind of steer teaches the model your judgment, and the same judgment shows up in every later prompt.

Save the brief somewhere you can paste it from, such as Notion AI, and plan to paste it at the top of every prompt from here on. ChatGPT is free to start, with Go at $8 per month and Plus at $20 per month removing the usage caps free plans apply at peak hours, so this step costs nothing.

This step collects the facts, statistics, and anecdotes the speech will quote, and it exists because the fastest way to lose a room is to say something the audience knows is wrong. Open Perplexity, switch to Pro Search, and enter this prompt with your topic filled in:

I am giving a 10 minute keynote to early stage startup founders on why most products fail at launch. Find me: three statistics on product failure rates from sources I can cite by name and year, one memorable example of a famous product failure and what it teaches, and one counterintuitive finding that will surprise this audience. For every fact give the source name, the year, and a link, and flag anything where sources disagree. Keep the total under 400 words.

Before you accept any number, click the citation and read the source yourself, because a speech is delivered live and there is no edit button in front of an audience. Keep only facts you would be comfortable defending if someone challenged them at the coffee break, and note the source name and year next to each one, because Step 3 will be instructed to use only these facts and nothing else.

Alongside the researched facts, write down two or three stories from your own experience that fit the topic, in rough notes at first. AI cannot invent your stories, and the speeches audiences remember are built on them. Perplexity has a free tier that covers this step comfortably, with Pro at $20 per month for heavier research use, so the cost here is also zero.

This step produces the complete first draft, and Claude is the tool for it because it sustains a spoken tone across long text better than any other model we track, which is why it holds 4.6/5 in our database. Create a new Project, paste in the brief from Step 1 and the evidence from Step 2, and use this prompt:

Using the speech brief and the evidence I provided, write a complete draft of my speech. Requirements: 1,400 words for a 10 minute slot at a natural pace of 140 words per minute. Open with the story angle I chose in the brief, then give a one sentence roadmap, develop exactly three points where each point opens with a short claim and lands with a story or a cited fact, and close with a call to action the audience can act on this week. Write for the ear: sentences under 20 words, no jargon, plain words over formal ones, and one deliberate repetition per section for rhythm. Use only the facts I provided and name the source for each. Mark the sections in brackets as you go.

Read the draft once without editing, and mark three things only: where your attention drifted, where you would stumble reading aloud, and which line you want to keep word for word. Then send one revision prompt, for example: the second point drags in the middle, cut 80 words there and replace the statistic with a concrete example from my evidence list. One targeted revision round beats five rounds of scattered tweaks, and it is how professional speechwriters work too.

If the first draft comes back too formal, which is the most common single complaint, do not edit it line by line; instead send one global prompt: rewrite the whole draft at a ninth grade reading level, cutting every sentence over 20 words in half. Fixing register in one pass, then polishing locally in Step 4, produces a more consistent tone than mixing the two jobs, and it takes about two minutes.

Claude is free to start, with Pro at $20 per month, or $17 per month billed yearly, for heavier drafting, so the core creative step of this workflow also runs on the free tier if you work within its limits.

This step turns a strong draft into a clean script, and it uses two tools for two different jobs: Grammarly for correctness and tone across the whole document, and Wordtune for sentence level rewrites where the rhythm feels off. First, paste the full draft into Grammarly, set the goal tone to confident and the domain to general, and work through its suggestions in one pass. Pay most attention to the clarity suggestions, because long nested sentences that look fine on the page are the first thing that breaks when spoken aloud, and cut every sentence over 25 words into two.

Next, move to Wordtune for the five to ten sentences you are least happy with, and rewrite each one two or three ways. Spoken language rewards short declaratives, and seeing three alternatives side by side is faster than staring at one sentence trying to fix it in your head. Use this pattern if you want the rewrite to stay consistent with the speech voice:

Rewrite this sentence for a spoken speech, keeping it under 18 words and sounding conversational rather than written: [paste the sentence]. Give me three versions: one plainer, one warmer, and one punchier.

Finish with a read-aloud pass in your own voice, even before the AI rehearsal in Step 5, and restore two or three phrases you actually say in everyday conversation, because that is what keeps the speech sounding like you rather than like a polished stranger. Grammarly is free with Pro at $12 per month billed yearly, and Wordtune is free with Plus at $6.99 per month, or $4.89 billed yearly, so this step stays inside the free tier if you need it to.

This step is the one most speakers skip and most audiences notice, and AI makes it cheap: paste the finished script into ElevenLabs, pick a natural voice, and generate a read-back you can listen to at normal speed with a timer running. Three things to mark while you listen. First, the total time, because anything more than 5 percent over your slot needs a cut today, not on stage. Second, every line where the synthetic voice trips, because a machine stumble is a reliable preview of a human one. Third, the sections where nothing changes for 90 seconds, because those are where a live audience starts checking a phone.

Then do at least two live passes yourself, standing up, and compare your spoken time against the AI read-back. Most speakers run faster than they expect, and the difference tells you how much margin to leave for laughter and applause. Use this prompt if you want the AI to help you cut to the clock:

My speech runs 11 minutes but my slot is 10. Cut it to 1,400 words without losing the three main points or the closing call to action. Prefer tightening sentences over deleting whole sections, and list every cut you made in a summary table so I can approve each one.

ElevenLabs is rated 4.6/5 in our database and free to start, with Starter at $6 per month and Creator at $22 per month for more characters and professional voices. Speechify at 4.5/5 is the mobile alternative at $29 per month, or $139 per year, if you want to rehearse during a commute, and both run the timing math for you: 10 minutes at 140 words per minute. Rehearse the opening and the closing twice as often as the middle, because those are the two moments where speaker and audience are both most nervous, and the AI read-back makes a six pass habit cheap enough to actually keep.

This step is optional for toasts and essential for keynotes, and it exists to support the speech rather than repeat it. Open Gamma, choose generate from text, and paste your finished script into this prompt:

Create a 7 slide deck to support a 10 minute speech. The speech script is below. Slide one is the opening image and the takeaway in five words or fewer, then one slide per main point with a single visual idea and no more than 15 words each, a slide with the key statistic from the script displayed large, and a closing slide with the call to action. Keep the deck minimal: the audience listens to me, the slides only anchor the points.

[paste your finished speech script here]

The discipline that keeps slides useful is the 15 word ceiling per slide, because slides that carry sentences steal attention from the speaker, and the audience reads faster than you talk. Use the one click restyle feature to test two or three themes before committing, and export to PDF as a backup for the venue laptop. Gamma earns 4.6/5 in our database as the fastest prompt-to-deck tool, it is free to try with 400 credits, and Plus at $9 per month covers regular speaking if slides become a habit.

If the speech needs charts instead of generated visuals, Canva AI at 4.4/5 handles branded graphics and the Brand Kit keeps colors consistent across future talks, with a free tier and Pro at $18 per month, or from $12 per month billed yearly. For a first deck, Gamma alone is enough.

The six step sequence stays the same for every occasion, but the emphasis moves, and knowing where to spend the extra hour is what separates a speech that fits from a speech that merely works. Four common types cover most situations, and each shifts the weight to different steps of this workflow.

• Wedding toasts and personal celebrations: spend the saved research hour on Step 1 instead, collecting three specific memories with the couple or family, and ask Claude for the classic arc of warm to funny to heartfelt. Keep it to 3 minutes, which is about 450 words, because toasts run long when nerves take over.
• Business keynotes and conference talks: this is the full workflow as written, with extra time in Step 2 on citable statistics and in Step 6 on the deck, because a projected statistic displayed large in Gamma does more work than three sentences of narration.
• Investor and sales pitches: weight Step 1 toward the single ask and Step 2 toward market numbers, and rehearse Step 5 twice as much, because a 5 minute pitch that runs 6 minutes reads as poor judgment to the room. Cut the deck to 8 slides hard.
• Academic and competition speeches: check the rules first, as covered in the FAQ, then weight Step 2 heavily, because judged speeches reward cited evidence; ask for a formal version and a conversational version of the draft, and choose after hearing both aloud in Step 5.
• Eulogies and memorials: the AI handles structure and gentle polish only, while every memory comes verbatim from family. Ask the model for an outline rather than a draft, write the stories in your own words, and use the ElevenLabs read-back to check the text survives a shaky voice, which matters more than any polish.

The pattern across all four: the more personal the occasion, the more the workflow belongs to you and the less to the tools, and the more professional, the more the research and rehearsal steps earn their hour.

These are the habits that separate an adequate AI-assisted speech from one that gets quoted back to you afterward, and each one costs minutes, not hours.

• Write for the ear with the 140 rule. Set 140 words per minute as the constant in every prompt, so a 5 minute slot is 700 words and an 8 minute toast is 1,120; the number in the prompt is what keeps the draft inside the clock instead of trimming in a panic the night before.
• Paste the brief into every prompt. The brief from Step 1 is the memory of the workflow, and pasting it at the top of each step keeps tone and audience consistent across tools, which is the difference between a stack and a workflow.
• Ask for three openings and test them aloud. Openings decide whether the room leans in, and generating three angles costs one prompt; read all three aloud in Step 5 and keep the one you can deliver without looking at the page.
• Generate a question drill before the big speech. Ask the AI to list the eight hardest questions a skeptical listener could ask after this speech, then answer each one aloud once; the confidence this buys is disproportionate to the 15 minutes it takes.
• Keep spoken paragraphs at 60 to 80 words. Long paragraphs hide long sentences, and long sentences are what a listener cannot follow; short blocks also make the cut-to-clock edit in Step 5 far easier.
• Mark the pauses in the script. Add a bracketed pause where a point should land, and the AI read-back plus your live passes will both respect it; a speech without designed pauses is a sentence without punctuation.
• Keep a swipe file of your best lines. Save the lines that landed, from your speeches and others, in Notion AI, and paste three into the drafting prompt; your second speech takes half the effort of your first.

Most AI-assisted speeches fail in one of five predictable ways, and every one of them is avoidable with a rule that fits in a sentence.

• Letting the AI invent facts. Models will produce confident statistics that do not exist, and a live audience has no edit button; the rule is that every number and quote in the speech traces to a source from Step 2 that you personally clicked and read, and the draft prompt already restricts Claude to supplied facts.
• Skipping the audience brief. A prompt that opens with write me a speech about resilience produces the same speech everyone has heard; the 20 minutes in Step 1 is the cheapest quality lever in the workflow, and skipping it is why AI speeches get called generic.
• Writing for the page instead of the ear. Drafts default to written rhythm, with long clauses and abstract nouns, and the fix is mechanical: sentences under 20 words, one idea per paragraph, and the read-aloud test in ElevenLabs before you commit to anything.
• Polishing until it sounds robotic. Perfect grammar everywhere reads as synthetic, so after the Grammarly pass, deliberately restore two or three phrases you actually say, including a fragment or two; audiences trust voices that sound lived in.
• Rehearsing the words but never the clock. A speech that is 15 percent long gets cut off mid-close, which undoes 30 minutes of good writing; run the timed read-back in Step 5 and cut to 95 percent of the slot, leaving room for laughter and for the slow first sentence.

Every tool in this workflow, with the step it owns, the price where the free tier ends, and what the free plan includes. Ratings are from the AITokenHub database, where these tools rank among the strongest in their categories.

Three patterns in the table are worth noting. The entire stack is affordable because every entry has a real free plan, which means the workflow can be tested end to end before any money is spent. The paid entry points cluster between $6 and $29 per month, so upgrading one bottleneck at a time is a low stakes decision. And no single tool covers more than two steps, which is why speakers who get consistent results run three or four tools rather than hunting for one app that does everything badly.

The honest answer first: one important speech per year needs nothing beyond free tiers, because every step in this workflow runs on a free plan and the constraints are usage caps rather than missing features. The upgrade math only starts to matter when speeches become a habit, and then each upgrade buys something specific.

Claude Pro at $20 per month, or $17 per month billed yearly, buys the highest usage ceiling and priority access at peak hours, which matters if you draft weekly or work evenings when free capacity runs thin; it is the first upgrade for most frequent speakers. ElevenLabs Starter at $6 per month buys more characters per month for the rehearsal read-back, enough for several full length speeches. Grammarly Pro at $12 per month billed yearly buys the tone and full sentence rewrites that free covers only in previews. Gamma Plus at $9 per month buys unlimited AI credits past the initial 400, which is one or two decks.

Run the arithmetic and the full core stack at yearly billing rates comes to about $44 per month: Claude at $17, Grammarly at $12, ElevenLabs Starter at $6, and Gamma Plus at $9. Adding Wordtune Plus at $4.89 billed yearly brings the total to roughly $49. That is less than one hour of a human speechwriter, and the honest recommendation is to stay free until a cap actually slows you down, then upgrade the single tool that binds first.

If the speech is within a week, this schedule turns the workflow into two short sessions, and every step below links back to the section that details it.

• Hour one today: run Step 1 in ChatGPT to lock the brief, then Step 2 in Perplexity to gather cited facts, and write down the two personal stories the speech cannot do without.
• Hour two today: draft in Claude with the full prompt from Step 3, do one targeted revision round, and leave the draft alone overnight, because distance finds problems re-reading cannot.
• Hour three tomorrow: edit in Grammarly and Wordtune, then run the timed read-back in ElevenLabs and cut to 95 percent of the slot.
• Before the day: build the deck in Gamma if the speech needs one, do two full live passes standing up, and run the eight question drill from the pro tips once.

Three hours is not a compromise; it is the same process professionals use, with the typing removed. The single rule that governs the whole checklist: AI drafts, you decide. The tools compress the hours that used to make speeches expensive, and the three hours they leave you are best spent on the parts only you can do, which are the stories, the judgment, and the delivery.