I've been writing prompts for AI tools almost every day for the past two years. Mostly for decks, because that's what we do at Huanjian. You'd think I'd have it down by now. I don't. But I've picked up a few patterns that save me from going back and forth with the model for an hour.
The biggest mistake I see is people treating the prompt like a search query. You type "quarterly business review" and expect something useful. What you get is a generic template that could be about any company in any industry.
How to Write AI Prompts for Presentations That Actually Work
I used to think a prompt was just the instructions you type. Then my teammate pointed out that the best prompts are basically a small brief. They tell the model not just what to produce, but who it's for, what you're trying to achieve, and what you're willing to leave out.
Here's a concrete example. The weak prompt: "Create a pitch deck for a SaaS product." That's fine if you want a placeholder. But if you're building something real, you need to give the AI anchors.
Try this instead:
- Product: Huanjian, an AI presentation tool that converts prompts into slide decks
- Audience: startup founders who hate designing slides
- Goal: get them to try the free trial
- Tone: honest, a bit technical, no hype
- Structure: problem, how we solve it, why it's different, pricing
That's not a fancy prompt. It's just context. But when you feed it in, the model suddenly knows what to do. It won't give you "Innovation meets productivity" nonsense, because you've told it to avoid hype.
I've started writing prompts like I'm briefing a junior designer. Not dictating every line, but giving constraints and reasoning. That shift alone cut my rewrite cycles in half.
Treat the first output as a draft, not a delivery
Even with a good prompt, the first version is rarely right. That's not a failure of the AI. It's how working with a tool should feel. The trick is to turn your feedback into the next prompt instead of editing the output manually.
In my experience, the best follow-up prompts are short and specific. "Cut slide 3 to one sentence." "Make the tone less corporate." "Move pricing before the metrics slide." Each one is a little correction, and the model keeps the rest of the context intact.
This is where I see people give up. They expect the first output to be the final deck, and when it isn't, they assume the tool is useless. What actually works is treating the conversation like an iterative process. I've had prompts that took eight or nine rounds before they clicked.
Honestly, I'm still figuring out the right balance between over-prompting and under-prompting. If I give too much detail, the AI gets rigid. If I give too little, it goes generic. The sweet spot seems to be giving the "why" behind each request, not just the "what".
While building Huanjian, we've seen users get much better results when they think about the flow of the entire presentation before they type anything. You don't need a full outline, but knowing your main message and the one thing you want the audience to remember helps more than any template.
A final thought on wording. You don't need to be polite to the AI, but being specific about what you don't want is surprisingly useful. "No bullet points longer than ten words" works better than "make it concise".
That's about it for now. I'm going to try a different prompt structure for my next deck and see if it holds up.