The Exact Prompt I Use to Generate an Investor Deck
Image generated by AI

The Exact Prompt I Use to Generate an Investor Deck

I have generated four investor decks this year with the same prompt skeleton. Not because I am lazy - because a deck is 80% structure and 20% judgment, and the structure part is exactly what an AI does well. The judgment part is still mine. So here is the prompt, section by section, and what I always fix afterward.

Why I stopped writing decks slide by slide

Old habit: open a blank presentation, stare at slide 1, write “Problem,” get stuck, close the laptop. 15 years of client work taught me structure first, prose second. But applying that discipline by hand every time is slow, and I would drift - too much text on slide 4, no clear ask on slide 11.

So I write one long prompt instead. I give the AI everything it needs to make the structural decisions, then I edit the output. The prompt takes ten minutes to write once and I reuse the skeleton for every deck since.

Part one: company context

This is the part founders skip and it is the part that matters most. I paste in:

  • One paragraph on what the company does, in plain words, no jargon
  • The specific problem, with a number attached (“teams lose 6 hours a week to X”)
  • Who pays for it today and what they pay
  • Traction: revenue, users, or pilot customers, with real figures, not “strong growth”
  • The raise amount and what it is for

If I skip the numbers, the AI fills gaps with vague filler like “significant market opportunity.” Garbage in, garbage out applies to decks the same as code.

Part two: audience and tone

I tell it who reads this deck. A seed-stage angel round gets a different deck than a Series A pitch to a fund partner. I specify:

  • Stage of the raise
  • Whether this goes out cold (needs more explanation) or gets presented live (needs fewer words, I talk the rest)
  • Tone: direct, no hype words, numbers over adjectives (yes, I ask an AI to write like I write)

I explicitly ban “revolutionary,” “game-changing,” and “seamless.” Every founder deck I have read this year had at least one of those words on slide 1. Investors skim past them now.

Part three: structure constraints

This is the part that saves the most time. I do not let the AI decide slide count or order freely. I give it the shape:

  1. Cover
  2. Problem
  3. Solution
  4. Product (screenshot or diagram placeholder)
  5. Market size, bottom-up not top-down
  6. Business model
  7. Traction
  8. Go-to-market
  9. Competition, honest positioning not a quadrant chart nobody trusts
  10. Team
  11. Financials, 3-year view
  12. The ask

Twelve slides, not twenty. I say this directly in the prompt: “Twelve slides maximum. If a topic does not fit, cut it, do not compress two topics onto one slide.”

Part four: design constraints

Here is where most AI-generated decks fall apart, and it is worth naming the failure modes so you can prompt around them.

Over-explaining. Left alone, an AI writes six bullet points per slide, each one a full sentence. Investors do not read sentences on slides, they read three words and glance at you. So I add: “Maximum 5 words per bullet. Maximum 4 bullets per slide. No full sentences.”

Bad chart types. Ask for “market size” and you will get a pie chart. Market size is not proportional data, it is a stacked bar showing TAM, SAM, SOM, or nothing at all. I specify chart type per slide where a chart appears, rather than trust the model’s instinct.

Wall of text on the problem slide. The model wants to justify the problem with three paragraphs. One paragraph, one number, done. I say this twice in the prompt because it ignores it once.

None of this is exotic. It is the same discipline I would apply writing the deck myself, just written down so the AI applies it too.

What the output actually looks like

I write the whole thing as a plain text file - headers for slide breaks, dashes for bullets, no formatting beyond that. This is where Slaide comes in, the deck tool I built after getting tired of PowerPoint mangling my layouts every time I edited on a different laptop. A plain text file renders straight to a web deck, a PDF, and a PPTX if someone insists on that format. I have an Agent Skill and an MCP server so an AI can author decks directly in that format, which is how I generate the first draft in one pass instead of copy-pasting into a slide tool afterward.

One honest limitation: if you go back and heavily hand-design a deck in Slaide’s editor - custom positioning, layered images, precise typography - that does not survive the PPTX round trip cleanly. Fine for a text-and-chart investor deck. Not fine if you want pixel-perfect brand decks exported to PowerPoint.

The editor itself is free for anyone signed in. The only thing that costs money is the hosted AI generation, meaning if you bring your own agent (Claude, whatever you use) and generate the deck yourself the way I described above, it costs nothing. The core tool is Apache-2.0, so you can also just run it yourself.

What I always fix by hand

Every generated deck needs the same three fixes, without exception:

  • The traction slide undersells real numbers or oversells soft ones. I rewrite it with the actual figures every time.
  • The competition slide is too polite. AI models are trained to avoid saying anything harsh about anyone, including competitors. I go back in and make the positioning honest.
  • The ask slide is vague (“raising to accelerate growth”). I replace it with the actual number and three concrete uses of the money.

None of this takes long, maybe 20 minutes total. But it is the 20 minutes that separates a deck that reads like every other AI-generated deck from one that reads like a founder who knows their numbers cold.

The prompt, one more time

If you want the short version: give it company context with real numbers, tell it who reads the deck and in what setting, hand it a fixed 12-slide structure, and ban vague adjectives and pie charts. Then read every slide yourself before it goes anywhere near an investor’s inbox. The structure is a starting point, not a finished product.