I've been writing prompts for AI presentation tools almost every day for the past year. My team builds one, so I've had plenty of time to see what works and what produces slide decks so generic they could have been made by a fortune cookie. The biggest mistake I keep seeing is people treating the AI like a search engine. You don't ask it to "make a deck about Q3 growth" and hope for the best. You give it a job.
Give the AI a job, not a command
The difference is subtle but critical. A command is "write 5 slides about remote work." A job is "you are a product lead who needs to convince a skeptical VP that remote work doesn't hurt velocity. The VP cares about metrics and precedent. The tone should be calm but firm, and every slide should lead with data."
That second prompt isn't harder to write. It just requires you to think for thirty seconds about who is listening and what you actually want them to feel. I've had success with a simple formula: audience, goal, objection, tone. Those four inputs alone turn a generic deck into something that feels like it was made by a human who understands the room.
My teammate told me last week that he started adding a fake constraint to his prompts. Instead of saying "explain our pricing model," he writes "explain our pricing model to a smart 15-year-old who has never seen a SaaS contract." The results are somehow clearer for real audiences too. Forcing the AI to simplify creates structure that adults appreciate. I've stolen his idea.
Break the deck into pieces
Another thing I've learned the hard way: never ask for a full deck in one go. The output gets repetitive and the logic starts to crumble around slide five. I've started asking for an outline first. Then I pick the sections that need more depth and ask for those separately. It takes more chat turns but the final product is always better.
Presentations are not documents. They're a sequence of visual moments. So I tell the AI to think in pictures, not bullet points. Instead of asking for a list of market trends, I ask for a visual metaphor for each trend. One time I got "we're all squeezing through a straw that's getting narrower" for a supply chain presentation. That stuck.
Also — and this might sound silly — clarify the story arc. Ask the AI to structure the deck like a narrative: setup, conflict, resolution. It doesn't work for every topic, but when it does, the flow becomes obvious. Your audience remembers stories, not slide headings.
Iterate like you're negotiating
The first draft from any AI tool is rarely good. It's a starting point. The skill is in the revision prompts. I've found that short, specific corrections work better than vague ones. "This slide sounds timid, make it direct" or "remove the word 'utilize' everywhere" or "the third slide's conclusion doesn't follow from the evidence, fix that."
While building Huanjian, we noticed something about how people use our tool. The ones who got the best presentations didn't write long, detailed prompts from the start. They wrote short prompts, looked at the output, then wrote another short prompt to adjust. Iterative, almost like talking to a designer. The long-prompt folks ended up with overstuffed, rigid decks that were painful to edit.
Honestly, I'm still figuring out the exact boundaries of what these models can do. Sometimes the AI ignores your constraints and you have to repeat yourself. Sometimes it surprises you with a metaphor you wouldn't have thought of. The trick is to treat it as a collaborator that lacks memory and taste, but has infinite patience. You bring the judgment, it brings the heavy lifting.
One last thing: don't be afraid to ask the AI to rewrite a particular slide from scratch with a different angle. I've seen people accept a mediocre slide just because it's already there. That's a mistake. The cost of regenerating is near zero. You're the product manager of your own presentation. Act like it.