I've been a product manager for about a decade. I've seen Gantt charts come back in style, watched Notion eat the wiki world, and sat through enough "synergy" decks to last several lifetimes. Lately, I've been deep in building an AI presentation tool called Huanjian. And the single biggest thing I've learned is that the quality of the slide deck directly mirrors the quality of the prompt. Most people blame the AI when the output looks generic. Usually, the prompt is the problem.

1. Stop telling the AI to "make it pop"

I see this constantly. Someone pastes a raw chunk of a Notion doc into ChatGPT and says "turn this into a presentation." What comes out is usually a bland, bullet-pointed mess. It's technically a presentation, but it has no soul, no focus, no point of view.

The fix is deceptively simple: treat the AI like a junior designer who has zero context about your business. Because that's essentially what it is. You need to feed it constraints. A teammate of mine asked for a "modern, sleek" deck last week. The AI gave her a black background with neon green text. Technically modern, completely unreadable on a projector. She had to specify: "light mode, company font, no more than 3 colors."

What worked for me is framing prompts like a brief. Not "create a deck about Product X," but "Create a 5-slide deck for Product X. The audience is skeptical engineers. They don't care about ROI, they care about architecture. Keep it minimal, use system design diagrams." When I started adding constraints like "who is the audience" and "what is the single goal," the outputs suddenly became usable. It's not magic. It's just recognizing that AI has an infinite possibility space, and your job is to narrow it down hard.

2. Iterate like you're coding, not like you're writing

I don't know anyone who writes perfect code on the first try. But for some reason, when people use AI, they expect a perfect deck on the first prompt. That's not how it works. Honestly, I'm still figuring out the exact syntax for this, but I've found a rhythm that works.

My process usually looks like this:

  • Prompt v1: "Here's a rough outline, make slides." The output is bloated and misses my point.
  • Prompt v2: "Remove slides X and Y, change the tone to skeptical." Output is okay, but slide 3 is still useless.
  • Prompt v3: "On slide 3, don't list features, list benefits. Add a quote from our CTO here." Output is genuinely good.
The temptation is to start from scratch. Don't. Iterate.

While building Huanjian, we noticed that the best results came when people could edit the prompt after seeing the first draft. So we designed the interface around this loop—generate, tweak, regenerate. The tool doesn't matter much. What matters is that you treat the AI as an engine for drafts, not a publisher of final decks