The question showed up in one of my group chats again last week: does anyone actually know of a good AI agent for pitch decks, because every one people try seems to disappoint them. She'd already landed on the workaround most people find on their own eventually — sitting inside a chat window, updating one slide at a time by hand. Still faster than opening Canva and starting from a blank page. Just brutally slow for something that's supposed to be the whole point of using AI in the first place.
I get some version of this question weekly at this point. So here's the real answer — the one I actually gave her, and the one I'm tired of retyping into DMs.
Why the generators keep letting people down
Gamma is genuinely the strongest of the bunch right now — full deck in under a minute, and the conversational editing loop is the smoothest in the category. It just exports to PowerPoint badly, which matters the moment anyone downstream needs an actual .pptx file. Beautiful.ai goes the other way — a 4.7 out of 5 on G2 and real design guardrails — but the writing it generates needs more editing than Gamma's. And Tome, which was the name everyone said first two years ago, is basically the cautionary tale of the category now: $81.6 million raised, 25 million users, and the presentation product got quietly sold off and shut down in 2025.
None of that is really the core problem, though. The problem is what these tools are optimized for. They're built to look impressive in a 60-second demo — bold layouts, a lot of icons, a gradient in the corner, the visual language that's already instantly recognizable as "someone ran this through an AI generator." What they're not built to do is sound and look like you — your actual colors, your actual argument, your voice in the room. A pitch deck isn't a template with a logo dropped in. It's your case, made visual. That part still has to come from someone who knows the case.
What I actually do now
Most of the time, I build the deck in Claude Code using a custom design skill, and I have it output HTML instead of slides. I host it as a link, or turn it into a PDF if someone needs a file. It's a faster loop than fighting with a slide tool — I can say "tighten this section" or "the second half is losing the thread" in plain language and watch it happen — and it takes way less back-and-forth, and way fewer tokens, than iterating on a PPTX file slide by slide. For the decks that genuinely need to live in PowerPoint — a client asks for it, someone's presenting from a laptop that isn't mine — I'll still run Claude Code with the same brand skill and have it output a .pptx directly.
- Claude Code + a design skill — builds the deck, usually as HTML, for the fastest iteration and the lowest token cost
- The brand skill — the thing that keeps every output on-brand instead of generic
- Claude Code + brand skill → PPT — for the decks that have to live in an actual PowerPoint file
- Nano Banana (Gemini's image model) via /banana-claude — generates and drops in images instead of a stock-photo hunt
- Canva MCP — just connected it, haven't tested it on a real deck yet; my guess is brand skill + Canva MCP ends up being the strongest combination going
Canva's still a good option, especially if design flexibility matters more to you than speed. I just get the best results, and the most efficient token usage, out of HTML. Your mileage may vary depending on what you're actually building.
The part that actually matters: the brand skill has to be specific
A vague brand skill produces vague output. Tell Claude to "use our brand colors, keep it modern," and it's going to improvise — and improvising is exactly how you land back on the generic AI-deck look you were trying to avoid in the first place.
Write it like you're training a junior designer on their first day, because that's functionally what it is. Exact hex codes, not color names — "brand pink" means nothing, #E8185A means something. Every font in the system named down to family, weight, and where each one lives — what carries a headline versus what carries a caption. Spacing and layout rules, not "clean and modern" — how much white space, how text and images relate to each other, what a title slide looks like versus what a data slide looks like. Logo rules: minimum size, clear space, what never happens to it. And the two categories almost every brand skill I've seen skips entirely — the actual voice and tone, written as real example sentences instead of adjectives like "confident" and "approachable," and an explicit list of anti-patterns: the phrases, layouts, and habits you never want to see, spelled out so the model isn't left guessing where the line is.
The more specific the skill, the less "AI" the output looks. Not because the model got smarter. Because it stopped improvising and started following your actual system.
If Claude Code is still new territory and you want a plain-language starting point before you build anything brand-related on top of it, I wrote this one first — worth ten minutes before any of this.