How I use AI to help me prepare talks

I’ve recently changed my workflow for preparing talks. I prepared 3 with this approach so far and I’m quite happy with the process so I decided to document it.

Inspired by Gary

I love Gary Bernhard’s How to prepare a talk. But it’s a lot!

tl;dr: improvise a bunch of times and the structure will crystallize. As a bonus you’ve now practiced a lot and are well prepared now.

I’ve tried this but it’s hard to find time (well, motivation really) to spend that much time on initial ad-lib runs. So I tried my own twist.

Recording

I’ll usually think about the structure a bit. But nothing fancy, not taking any notes. And then I’ll start out the same as Gary: pace around my room and just talk to no-one. It’s….bad. But it’s a great way to just get started and get one full take done quickly. No writers block of staring at an empty page.

But I’ll do two alterations here. First: I’ll do “note to self” breaks. Just stop talking about the content, say “note to self” and dump my thoughts. This will stuff like “I should actually explain X in more detail earlier” or “a Y diagram here on slides”.

Second: I’ll record myself. No intention of ever listening to that but I’ll run it through a transcription model (usually whisper on my laptop) to get “subtitles” for my take. So timestamped transcriptions.

Notes to self

Then I’ll run this through an LLM with the prompt to extract all notes to self. This will go to my scratch pad / todo list. Usually I’ll alter the content quite a bit when writing it down. But the idea is that I don’t stop talking while doing my take to type something down and instead just dictate everything. And at least for me it works wonders to keep me in the zone.

Early feedback

Sometimes I’ll do multiple takes first, sometime I’ll do this on the first one already: I’ll run the transcript through an LLM and ask for feedback, specifically instructing it to skip delivery, phrasing etc but focus on the content, what would need more explanation, more clarity, or less verbosity. I find it a great way to get some quick inputs very early without subjecting people to my very early bad takes or myself to the stress of doing early improvisations with live audience.

Extracting structure

After I get to a decent take I’ll run that transcript through an LLM and prompt it to extract the outline of slides. It usually does an okay job of getting started. Again it’s a tool to defeat the empty page. Then I’ll look at that outline and I’ll do my slides around that. I’ll use some structure, and I’ll deviate. The key thing is that I’m in control and I use my taste to decide on the content. I think it’s important that the content I present (both the talking and the slides) is actually mine and not generated.

Practice

And then keep going until very comfortable. You just cannot beat the time investment. But I think AI tooling helps get my lazy self over the initial hump quicker.

Last modified on 2025-12-10