Given the interest I'm sharing the slides for my Wednesday talk at NVIDIA as an example. The slide deck was built upon a 174-line mental schema, with quarto + claude code (opus5.5 medium) + custom translation rules.
slides: https://tianshili.me/assets/slides/nvidia-contextual-boundaries/
mental schema: https://gist.github.com/i7mist/7446a8c6581715b2982efb88237bedcf
rules: https://gist.github.com/i7mist/813bfda0816c4d16e3007313a99b2136
You can read the mental schema along with the slides, or just read the schema itself! To me the schema itself describes my thought flow even more clearly, but it might also be the natural bias because I wrote it. Curious to know how you feel if you've read both versions.
One thing I like is that it gives me a good balance between delegation and direct manipulation. It has given me the precise control I need over the content, while abstracting away a lot of details that are less important and enjoyable yet took most of my time in my past slide making experiences. (this can be personal and I know there are people who enjoy working their visual design magic)
I do want to share some caveats: 1) the agent also generated speaker notes which are really bad (press S to see) 😅 I barely use speaker notes so have low incentives to fix them now. 2) I'm hoping this to be modular and reproducible, so the schema functions like code and the agent is the "compiler" that compiles the code into the artifact. But since the slide making process is very iterative which means I made many changes after the initial version by editing the schema and asking the agent to update it based on the diff, I'm not sure if it can be fully reproducible just with the final schema without all the editing history. It could be a cool project to explore when I run my course again and update all the slides next year.