Reverse-Engineering Slides for a Talk That Had None

I was going back through my AI Engineer World’s Fair 2026 watch list and re-watched Garry Tan’s “The New Physics of Business” keynote. About four minutes in, he says:

“Sorry, there’s no slides. I have no slides.”

No deck, the whole time. Just Garry, a mic, and 20 minutes of dense material about AI-native orgs, skill files, and “company brains.” I thought it’d be a fun experiment to see if Claude Code could reverse-engineer a slide deck purely from the talk itself — no slides to work from, just the words.

How it went

I pointed Claude Code (Sonnet 5) at the YouTube link and asked it to reconstruct slides. My prompt was:

Gary tan’s talk doesn’t have slides even though I think it should, can you reverse engineer some slides for his talk? https://www.youtube.com/watch?v=eBUyTS7SzV4

It:

1. Opened the video in a browser (via the Claude Code browser MCP tools), expanded the description, and pulled YouTube’s auto-generated transcript panel rather than trying to guess at content from memory — every claim in the resulting deck is traceable back to a timestamp in the actual talk.
2. Cross-referenced the transcript against the video’s own chapter markers and “key takeaways” blurb to build an outline: the 400x productivity claim, the “skill file = employee” org mapping, latent vs. deterministic space, the “company brain” concept, and the closing “we can boil the ocean now” line.
3. Used its frontend-slides skill to generate a single self-contained HTML file — no build step, no dependencies, just one file with inline CSS/JS. I let it make its own creative call on visual style rather than going through the usual mood/preview picker, and it landed on a dark theme with YC-orange accents (Fraunces for the big serif statements, JetBrains Mono for kickers and timestamps) — a nod to the fact this is literally a company brain and markdown-employees talk.
4. Verified the result in an actual browser — checked title, stat, quote, and card-grid slide types for overflow and layout bugs before calling it done.

Every slide is tagged with the timestamp range it’s drawn from, and the title slide is upfront that this is a reconstruction, not Garry’s actual deck (because there wasn’t one).

The result

→ View the slides

It’s a normal HTML page — arrow keys, space, scroll, or swipe to move between slides, click the dots on the right to jump around.

Why this is interesting

The mechanically interesting part isn’t the slides themselves, it’s that the whole pipeline — watch a video, pull its transcript, structure the content, design and build a polished artifact, verify it renders correctly — ran without me touching a slide tool, a design tool, or a video player by hand. A few years ago this would’ve been: transcribe by hand (or pay for it), outline in a doc, open Keynote, pick a template, lay out 20 slides, export. Here it was one prompt and a follow-up to publish it.

That’s basically the exact thing Garry’s talk is about.