On AI slides

I promised to say a little more about my experiment with the presentation for my last talk, which I vibe-coded. My previous attempts had all run into the same problem: the slides looked too AI-generated, and it was still easier to make them by hand to express my thoughts. I've already written about solving the first problem by creating a skill with slide templates. But how the agent gathered content and fitted it into those templates is interesting too.

The talk was about several of my repositories and how I'd worked in them with agents over a long period of time. I wanted to reflect on the process myself and understand it better. As usual, we started with me talking into a microphone, throwing out random thoughts about what might be interesting and adding technical details about what to look for and where, in those repositories and my agent session histories. The agent gathered information and, alongside those artifacts, produced something like a first script. There were quite a few iterations. At some point Hermes joined the work on the repository for preparing the talk, and I started throwing a lot of ideas its way on the go.

When I ran out of thoughts and deadlines were getting close, I sent the agent off to make the first draft of the presentation. It came back with about 500 slides for a 20–30 minute talk. Even allowing for my style of having a separate slide for every change, that's a lot. I didn't know what to do with it. There was far too much slop, not enough visual material, and the emphasis was on completely the wrong things.

Then I had a wonderful idea. I started recording my screen in OBS and spent several hours clicking through the slides, talking into the microphone about everything I thought of each one. Literally: "The first slide is fine. On the second slide, add a link to the channel and the company name after an @ sign... On slide 17, let's put a video in the background and add a fade-in. On slide 19, let's slowly scroll the image from top to bottom... On all slides like number 25, let's remove the subtitle under the heading. For transitions between slides 47–49, let's have the images slide in smoothly. On slide 50, record a video of the website and insert it as a GIF. On slide 92, let's animate an audio waveform. On all slides with demo videos, like number 95, let's speed the video up 1.5 times and add a rewind animation at the end so it can play on a loop... Slides 137–145 could be shown as one list..." That's not counting the ten-minute monologues where an idea would pop into my head about how to gather some more interesting information that wasn't in the script.

Then, hoarse from endless audio review, you take the video recording, give it back to the agent and say: "Transcribe it with timestamps," and "If I forgot to say a slide number somewhere, look at the video frame at that moment to see what I was looking at when I said it." You leave for an hour. When you come back, most of the work is done the way you wanted. You sit down with OBS for a second, shorter pass, and everything gets close to perfect. After that it's easier to finish off the small stuff with targeted prompts.

The agent also acts as a supervisor. It won't let you forget to record a video for a slide you mentioned somewhere and left as a placeholder. It'll remind you about slides marked "draft". It'll catch silly mistakes: things misaligned on a slide, wrong spacing, fonts or colors, or a stray flowchart line left behind after copying a slide (based on real events). I work with complete slides first, without revealing elements gradually. At the end, I split them up with a single prompt and review the final result.

Most importantly, there are things I would never even have tried to do by hand. And the more you do, the more ideas you get. Your mind starts coming up with things you'd have been afraid to consider before. Highlight code however you like? Sure. Make a slide like a YouTube channel, with moving videos? Sure. Reveal a tricky table step by step? Easy. Break a video into wireframes to show its structure? No problem. No images or videos suffered from my struggles in other programs; the agent processed them itself.

A disclaimer: I've developed my approach over all the talks I've given and, I'd like to believe, I know how to build a narrative. Every idea ultimately comes from me; the agent executes it. An inexperienced speaker might put too much trust in agents, and the talk could turn into a pile of fluff and unnecessary details, with slides full of useless, context-free text and errors. My presentation looks simple, but there's a lot of work behind that simplicity.