How I Use AI to Prep for Talks

I'm currently prepping for a presentation to one of the largest audiences I'll speak to this year: the global HubSpot team (~9,000 HubSpotters).

We're hosting what we call GrowDAI -- a day-long internal conference where we spend time learning about AI. The content is by HubSpotters for HubSpotters. Yes, we love AI, but we love people even more and want them to learn and grow.

And the hardest part of these presentations? Figuring out what to say.

Most people think I use AI to write my talks. I don't. I use it to figure out what not to say, what's missing, and what might fall flat.

So today, I want to break down:

• How I actually use AI for speech prep (sparring partner, not writer)

• The virtual audience technique

• What I'm testing for GrowDAI specifically

—@dharmesh

What Most People Do vs. What I Do

Most people use AI as a writer. "Write me a presentation about X." "Make it inspiring." "Add a story."

That's not wrong, but it's not how I use it.

My most common use of AI, outside of agent coding, is as a sparring partner. Not a writer but an editor and a critic. The kind of person you'd want reading your draft before it goes anywhere near a stage.

I use it to find gaps in my writing like: "Okay, Dharmesh, you said this right here, but it's logically inconsistent with this thing you said earlier." Or: "You said this, here are the gaps in your reasoning. This doesn't quite make sense, or this was not clear enough, or this might be controversial -- and you don’t make a strong enough case."

For the heavy lifting on this kind of work, I lean on Claude (but ChatGPT or Gemini can work pretty well too). Claude is what I use for almost all of my coding (I've written about my coding setup before), and it turns out the same technology that finds bugs in code is pretty useful for finding bugs in arguments.

But here's the catch: The model on its own isn't enough. What makes the sparring partner actually good is the context I give it.

As I’ve said before, context is queen. But what does that actually mean?

I’ve built a personal context store that I call my “Second Brain”. It has documents, my emails, past talks, past writing, etc. I take the relevant parts and give it to Claude. This means it can catch inconsistencies across years of content -- not just within one draft. It knows what I've said before, what's worked with different audiences, and what hasn't.

For this particular audience of 9,000 HubSpotters, Claude might tell me: "Based on the data of what's worked for you in the past, this section probably won't land the way you think it will."

I treat it conversationally. Not because I attribute human qualities to it, but because it helps to interact with it like I might a smart human that’s well educated. That’s how it was trained to work.

The Virtual Audience Technique

Beyond critique, I create virtual audiences to test my ideas before I ever step on stage.

I describe my audience to AI with as much detail as I can -- who they are, what they care about, what their backgrounds are. Then I pose "what ifs" and see how they react.

It's like having a customer advisory board of 12 people with specific profiles, and I can put this idea in front of them and see how they react.

For example for GrowDAI, I'm creating virtual HubSpotters right now: engineering leaders who've been here 10+ years, new sales reps who joined in the past 6 months, marketers experimenting on the edges, support folks dealing with customer questions about AI every day.

Then I test parts of my talk against them. "How will this be received?" "Is this too technical?" "Does this actually help them do their jobs better?" And the best question: "How can I make this simpler?"

The AI simulates what the real-world response might look like. It's not perfect, but it's great for figuring out format, content balance, which controversial takes to include, and how much vision vs. practical "try this Monday" advice to give.

It can sometimes help brainstorm new ideas well too. I don’t usually end up using these, but it can often spark other ideas that do end up working.

Which leads me to my next segue: What would you want to learn about if you were at GrowDAI? What are you curious about? If you were moderating a fireside chat with me, what would you ask?

Even if you're not on the HubSpot team (yet), imagine that you are. What would make this worth your time?

(Note: While the presentation isn’t going to be public outside of internal HubSpotters, I’ll do my best to share the materials that landed best in this newsletter following the presentation)

How You Can Use This Yourself

You don't need a 9,000-person room to get value from any of this. The same playbook works for a sales pitch, a board update, an all-hands, or a tough customer call.

It comes down to these five steps:

1. Describe your audience to AI with as much detail as you can. (remember, context is queen)

2. Share your draft talking points or outline.

3. Ask AI to critique it as if it were your audience.

4. Create virtual panels.

5. Pressure-test the bold takes.

If you're going to say something bold, ask AI: "Is this controversial? If so, for what reason? What's the strongest counter-argument?"

The goal isn't to eliminate everything controversial. It's to know what you're walking into and whether it's worth it.

—Dharmesh (@dharmesh)