I've been writing prompts for slide decks almost every day for the past year. Partly because I work on an AI presentation tool, but mostly because starting from a blank deck is a pain. I've made enough mistakes that I'm starting to see a pattern.
Why most prompts fail
The most common prompt I see is something like "make a presentation about our new product." That reads like a search query, not a brief. The AI has to guess who the audience is, what you want them to think or do, and what order the information should go in. So it falls back on a generic structure: title, problem, solution, feature list, next steps. That's not wrong, but it's also not useful.
Last week my teammate asked the model to "create a deck about the benefits of remote work." We got twelve slides that could have been written by any HR blog in 2021. Nothing was factually incorrect, and nothing was specific to our team, our clients, or our situation. The problem wasn't the tool. It was the prompt.
What actually changes the output
After a lot of trial and error, I've found that three things consistently improve the results. The first is giving the AI a point of view. Instead of "benefits of remote work," try "argue that remote work increases productivity, but only when you invest in async communication." Now the AI has a thesis to defend, and the slides will have a spine.
The second is naming the audience explicitly. "Present to a CFO who cares about cost savings" produces different slides than "present to a team that's skeptical about change." The model, as far as I can tell, does respond to that. Not perfectly, but noticeably.
The third is constraining the structure. Tell it how many slides you want, or what sections you want to avoid. For example, "don't include a slide about our team" or "the last slide should be a concrete next step." Without constraints, the model defaults to the safest possible outline, which usually means three bullet points per slide.
While building Huanjian, we noticed that users who wrote longer prompts with these details got decks they actually used. The ones who typed one sentence got something they had to rewrite entirely. So we started designing the interface to prompt for that information before hitting generate. It seems obvious now, but it took us a while to see that the prompt is not the input — it's the first step of the work.
A prompt pattern that works for me
Here's a rough template I've been using. I don't follow it every time, but it beats staring at a blank input box.
- Audience: who are they, and what do they already know?
- Goal: what should the audience think or do after the deck?
- Thesis: one sentence that states your position.
- Slides: a rough sequence of sections, not necessarily titles.
- Exclusions: anything you don't want to see.
A concrete example. Instead of "present our quarterly results," I might write:
"The audience is our investors. They've seen last quarter's numbers and they're worried about growth. The goal is to convince them that the drop in Q2 was a one-time issue. My