You can tell an AI-generated deck from across the room. Same hero image, same three-column "key takeaways", same earnest gradient, same rounded card with the icon centred above the heading. The model isn't bad. The substrate is wrong.
What an AI slide generator can actually edit
Inside Tome or Gamma, the model is producing styled text inside a fixed schema. Title, body, image. Title, body, image. The schema is the floor and also the ceiling. If you want a slide that doesn't exist in the schema — a real chart, a custom diagram, a live demo, a control panel the presenter tunes mid-pitch — there is nowhere for the model to put it.
PowerPoint is worse. The model can write you XML, technically. But nothing in its training set encourages it to write good .pptx. There are no public repositories of beautiful PowerPoint files. There are millions of beautiful websites.
What Claude Code is actually good at
Claude Code is not a slide generator. It is an extremely good React engineer. It has read every Tailwind class. It has refactored every Next.js app on GitHub. Ask it to build you a 16:9 hero with a left-aligned headline, a stat block on the right, and a thin progress strip across the bottom — it does it in three seconds and it looks correct.
The mistake the slide-generator companies made was: instead of pointing the best web engineer in the world at the problem, they pointed an LLM at a worse abstraction and called it AI.
You can't refactor a .pptx
This is the test. Ask any of them to:
- Standardise every headline in the deck to a single type scale.
- Pull the same metric out of three slides into a shared component.
- Swap the chart on slide 7 to read from a different data source.
- Add a presenter-only debug panel to slide 11.
- Generate a second deck that reuses 60% of the slides from this one.
These are five-minute changes in a React codebase. They are unanswerable in a slide generator, because the unit of editing is "the slide," not "the component," and the slide is opaque to the model.
The thing they're optimising for is the wrong thing
The pitch for AI slide generators is "make a deck in 30 seconds." That's a useful demo. It is not a real workflow. Nobody serious shows up to a pitch with a 30-second deck. People iterate on the same deck for months. The deck is alive.
The right thing to optimise is: how cheap is the next iteration? How cheap is "show me the same deck but with the Why-Now slide as an animated diagram instead of three bullets"? How cheap is "rebuild the IM from the pitch deck — same content, denser, with risk and financials sections"?
In a React codebase, those edits are import-level. In a slide generator, they're a rewrite.
What to use instead
Skip the slide generator. Build a Next.js app. Each slide is a component. Each deck is an array of components. Present in fullscreen. Export with html2canvas-pro + jspdf. Ask Claude Code to refine the slides the same way you'd ask it to refine any other route.
That is what nopoint is. It's the scaffold so you don't have to build the viewer, the navigation, the export, the manager, the keyboard shortcuts. You start with the runtime and write your slides.
The "AI" part isn't the slide generator. The AI part is your normal coding agent doing what it's already best at — writing a website.
Read next
- A slideshow is a website with a PDF export — the first-principles claim
- Deck-as-code: the slide registry — one viewer, many decks