I spent last week building an investment deck for a client. The raw material was a pile of research reports. The output needed to be a branded PowerPoint that looked like it came from inside their firm, not from a random consultant with a Canva account.

If you've ever tried to get an LLM to spit out a polished, branded deck, you know how this usually goes. The content is fine. The formatting is a disaster.

Here's what I tried.

Attempt 1. I worked in Claude, pointed it at the folder of research reports, and asked it to build a PowerPoint. It did. Formatting was clean. No branding, but clean. Worked in Slides too.

Attempt 2. Same setup, but this time I asked it to apply the firm's brand guidelines. Logo, colors, type, the works. Individual pages started breaking. Misaligned text boxes. Charts that didn't render. A title slide that looked like it had been through a paper shredder.

Attempt 3. I opened the deck directly in PowerPoint and asked Claude to reformat it from there. Worse. The model could see the slides but couldn't reliably manipulate them. Every fix introduced a new problem.

Attempt 4. I gave up on starting from the deck. I went to Excel, built the charts and tables I needed there, linked them properly, and then asked Claude (from inside Excel) to wrap a branded PowerPoint around the data. It worked. The brand skill fired. The pages held together. The deck looked like the firm. The Claude agent in Excel managed the Claude Agent in Powerpoint.

Here's the lesson, and it matters more for investors than I think people realize.

(1) The LLM does its best work closest to structured data, not closest to the artifact you want.

We keep trying to AI-ify the final output. The deck. The memo. The letter. That's the wrong end of the pipe. The model is being asked to do four things at once: read your source content, apply your brand, lay out your slides, and render them correctly. Something always breaks.

When the data lives somewhere structured (Excel, a database, a clean folder of reports) and the LLM builds the container around it, the work holds. You're letting the model do what it's actually good at: organizing and presenting structured input. You're not asking it to be a graphic designer working in a medium it can barely see.

(2) Formatting matters.

Claude broke when it needed to work with formatting it didn't understand. Logos in EMF not PNG. Graphs it doesn't know how to create or edit.

I'm rebuilding two more client workflows this way now. Charts and data in Excel, deck assembled on top. Will report back.

Alex

PS If you've found a deck-building workflow that solves the formatting problem differently, reply and tell me. I want to hear it.