A dark slide headed 'Let AI make the deck. For once, trust the result.' from the YC demo-day example deck.
From the YC demo-day example deck, rendered with slaide.

Using AI Agents to Build and Review Slide Decks

By · · Updated

You cannot hand a model a .pptx and ask it to fix slide 12. Not usefully. The model would have to either drive a GUI, which is slow and brittle, or edit raw OOXML, which is a format nobody should be editing by hand let alone asking a language model to patch blind. Either way you’re not getting a reliable result.

Hand it a text file instead and the job is just text editing, which is the thing these models are actually good at.

What’s actually wired up

Slaide ships an MCP server and an agent skill. The skill supports Claude Code, Codex, and Gemini. Any client that speaks MCP can use the server. The Slaide core is Apache-2.0. Run slaide install and it sets up both.

Here is the kind of prompt that works:

Write a five-slide deck summarizing this quarter’s numbers: MRR growth, churn, new customers, the two product changes we released, and next quarter’s one priority. Keep it to one idea per slide. Use the aurora theme.

What comes back is a .slaide file: Markdown for the slides, short lines under each. You render it, look at the PDF, and ask for fixes: a slide that’s too dense, a number that needs a source, a headline that states the obvious instead of the point. The agent skill automates part of this. It runs slaide validate --strict, shoots the slides into one image, reads that image, and fixes clipped text and weak contrast.

slaide export quarterly.slaide --pdf quarterly.pdf

That loop is prompt, render, look, adjust. I have not timed it against typing the deck into PowerPoint by hand, so I make no speed claim.

Review works the same way

Because the deck is text, “review this deck” is a normal code review prompt, not a special case. I can ask an agent to check a .slaide file against a style guide, flag any slide with more than three lines, or confirm every number in the deck has a citation in the notes. It reads the file and does not need to render anything first. For words and numbers, that beats asking a model to review a PDF or slide images, where it infers content from pixels. For layout problems, rendering still matters. That is why the skill also looks at a rendered image.

The honest limit

The agent writes text. It does not design. If you ask it for a “clean, professional deck” it will write clean, professional sentences and apply whichever theme you told it to use. It will not invent a new layout, choose typography, or make a judgment call about whitespace, because none of that lives in the text file it’s editing. A theme is a YAML file that names fonts, colors, and layouts. The agent picks from those names. That gives the model a small, well-defined surface to work in instead of an open-ended design problem.

That’s a feature for the decks this is aimed at, status reports, customer summaries, internal updates, anything where the content is the deliverable. It’s a real limit for anything where the design is the deliverable. Don’t ask an agent to build a keynote that needs a studio look. Ask it to write the ten sentences that are actually in your head, correctly, in the right order, and let the renderer handle the part that doesn’t need a design opinion.

My view: this fits decks where you know what you want to say and the friction is typing it into a GUI. The agent does not make your decisions. It saves you the ribbon menu.