I've made a lot of bad decks with AI. They all failed the same way: I asked for a presentation about something instead of a presentation that does something. The prompt was a topic, and the model handed back a topic-shaped pile of slides.

Here's what changed.

Say what the deck has to do

"Make a presentation about our Q3 pricing changes" is a subject line, not a brief. The model has no idea whether this is a five-minute standup update or a 40-slide sales enablement deck. So it picks something in the middle, and the middle is always wrong.

What works for me is writing the prompt the way I'd brief a freelancer who's never met the team. Who's in the room. What they already believe. What I want them to think or do when it's over. How long I have.

Something like: "There are 12 people in this meeting. Most of them think our pricing is too low and want to raise it. I'm going to argue that the problem isn't the number, it's that we've never tested a mid-tier. I have 10 minutes. Build a deck that gets them to agree to a two-week test, not a full repricing."

That's a completely different output. Not because the model got smarter, but because there's now a decision to aim at.

Paste your raw material instead of describing it

The laziest version of me still writes prompts like "include data on customer churn." The model then invents plausible-looking numbers, which is worse than no numbers at all.

When I paste the actual thing — the messy meeting notes, the spreadsheet dump, the Slack thread where the argument happened — the output gets specific in a way adjectives can't buy. Last month a teammate handed me a 900-word transcript of a customer call and asked for a deck. I dropped the whole thing in and added one line: "Every claim on a slide has to trace back to something in this transcript. If it isn't in here, leave it out."

That instruction mattered more than the length limit, the tone, or the "make it professional" stuff people usually pile on. Constraints about sourcing beat constraints about style.

One caveat: I'm still figuring out how much context is too much. Past a few thousand words the model starts treating earlier details as less important, at least in my experience. If your source material is a 40-page doc, trim it first, or feed it in chunks and ask for an outline before slides.

Ask for structure first, then write slides

One-shot prompts produce decks that read like essays with bullet points. The fix is boring: two steps.

Step one, ask for an argument outline. Six lines, one per section, each line stating what that section is supposed to convince the audience of. No slide content yet. Read it. This is where you catch the deck that has no spine, before you've spent twenty minutes telling the model to shorten sentences.

Step two, only after the outline is right, ask for slides. Then constrain hard:

  • One idea per slide. If a slide needs two headlines, it's two slides.
  • Headlines should be claims — "Churn is flat, revenue isn't" — not labels like "Churn Analysis."
  • Give a slide count. Say 12 out loud. Models will happily give you 30.
  • Tell it what to leave out. "No methodology slide, nobody in the room cares."

The headline-as-claim rule is the one I'd keep if I could only keep one. It forces the model to decide what each slide is arguing, and the filler slides have nowhere to hide.

Small things that help

Ask for three alternate titles instead of one. Titles are cheap to generate and weirdly revealing. If all three are boring, the deck probably is too.

Ask what's missing: "What would a skeptical VP push back on here?" I don't always take the answer, but it surfaces holes I'd have walked into live.

Skip words like "engaging" and "compelling." They don't correspond to anything the model can act on. Describe what the audience does with the information instead.

While building Huanjian, our AI presentation tool, we spent a lot of time on this, mostly because we kept watching people paste a one-line topic and then blame the output. The tool matters less than the brief. A good prompt on a mediocre tool beats a vague prompt on the best one, and I say that as someone who'd rather it weren't true.

My own prompts are still too long. I write them like I'm negotiating a contract. But the decks come out closer to usable now, and that's mostly because I stopped describing the presentation and started describing the room.