The Prompt Formula: How to Ask AI for Slides That Actually Work

September 16, 2026 · 14 min read

The gap between "make me slides about X" and a deck you'd actually present is almost entirely the prompt. A vague prompt makes the AI aim at the average of everything ever written about X: generic advice, invented examples, an agenda slide. The formula below has seven parts, takes two minutes to fill once you've written a brief, and works in any chat AI or in the "advanced prompt" field of slide generators. Think of it as a checklist disguised as a prompt — each line closes a specific gap the AI would otherwise fill with its own assumptions.

The formula

[ROLE]     You are a presentation designer writing for a [industry] audience.
[AUDIENCE] The audience is [who]. They already know [what]. They care about [what].
[GOAL]     After the talk they should [decision / action / belief].
[INPUT]    Base it only on this material: [paste notes, doc, data].
[FORMAT]   [N] slides. One idea per slide. Bullets under 12 words.
           Titles are takeaway sentences. Speaker notes on every slide.
[TONE]     [direct, numbers first / warm, plain language / ...]
[AVOID]    No agenda or thank-you slides. No generic stock advice.
           No invented statistics — mark missing data as [NEED DATA].

Diagnose before you fix

Bad AI decks are predictable, which is good news: each symptom points to exactly one missing part of the formula. Run this table in reverse against the last deck that disappointed you, and you'll know which line to add:

SymptomMissing part
A deck that could belong to any company on earthAUDIENCE
Plausible statistics nobody can sourceINPUT
Eight bullets of fifteen words on every slideFORMAT
Agenda slide, "Introduction", "Thank you!"AVOID
Covers the topic but takes no positionGOAL
Wrong register — too formal, too salesy, too chattyTONE
Explains basics your listeners wrote the book onAUDIENCE (the "already know" clause)

One missing line produces one failure mode, so you rarely need to fix more than two things at once. That's the practical advantage of a formula over "write better prompts": it turns a vague craft problem into a checklist.

Part by part, with the reasons

ROLE — set the profession, set the bar

"You are a presentation designer writing for a [industry] audience" mostly controls vocabulary and what the AI considers a good slide: a designer's outline has takeaways and hierarchy; an assistant's outline has topics and bullets. This is the least important part — with everything else filled in, the deck barely changes if you drop it — but it costs one line, so keep it. Where it genuinely matters: specialist contexts (medical, legal, financial, deeply technical), where the wrong register makes the whole deck read as amateur hour to the one person in the room who knows better.

AUDIENCE — the heavy lifter

Three clauses: who they are, what they already know, what they care about. The middle one is the most skipped and the most valuable, because it tells the AI what not to explain — and AI's default is to explain everything, which is how senior rooms end up watching a definitions slide.

The fastest way to feel why this matters: one topic, three audiences. Say the company is changing its pricing model:

Same facts, three different decks, and the difference was one line in the prompt. If you only ever add one part of this formula, add the audience.

GOAL — one sentence, one verb

"After the talk they should approve the migration plan / believe the churn diagnosis / book the pilot." A verb, not a vibe. The goal does two jobs: it decides the last slide (the deck ends on what the room is supposed to do), and it acts as an inclusion test you apply during the edit — every slide must move the room toward the goal, and slides that don't get cut. The AI can't run that test for you unless the goal is in the prompt, because from its side, "covers the topic well" and "wins the decision" both look like success.

INPUT — the difference between grounded and invented

Paste your material — notes, the document, the numbers — and add the phrase "Base it only on this material." Without input, the AI will invent plausible examples and statistics to fill the silence, fluently and without marking them. This is the single most dangerous failure mode in AI decks, and the fix costs nothing but a paste.

Two practical notes. First, curated input beats volume: for a 10-slide deck, one page of notes that states the real numbers and the position outperforms ten pages of raw documents, because the AI spends its attention on selection instead of extraction. Second, placeholder discipline: anything the material doesn't answer must appear as [NEED DATA] in the output, never smoothed over. A deck with three visible [NEED DATA] markers is professional; a deck with three invented numbers is a liability. You fill the markers yourself before the deck meets a room.

FORMAT — the constraints that visibly change output

Format lines do more for the final result than any design theme, because they fight the AI's default behavior — summarize everything, explain everything, hedge everything. The ones that earn their place:

TONE — three words that settle arguments

Tone lines that actually work are concrete: "direct, numbers first, no marketing language", "warm, plain language, explain jargon on first use", "technical, precise, costs and risks stated plainly". A useful trick inside the tone line is the banned-phrase list — "never use: synergy, seamless, leverage, in today's fast-paced world". Every office has its own words that make a deck sound like parody; name them once and the AI avoids them everywhere, which is more reliable than fixing them slide by slide.

AVOID — pre-ban the defaults

AIs don't infer taste. The agenda slide, the "Thank you!" slide, the triple summary at the end, the invented statistic, the opening line about how the industry is changing — these are defaults, and defaults appear unless banned. The avoid-list is three lines that save twenty minutes of deletion later. Add to it every time a default annoys you; the list only grows and never misfires.

Same topic, two prompts, different decks

Bad: Make me a presentation about our Q3 marketing results. You will get: Introduction, Agenda, Overview of Q3, Social Media Performance, SEO Performance, Email Performance, Key Learnings, Next Steps, Thank You. Every slide is a chapter heading, nothing is a claim, and the deck would survive any company swapping the logo.

Good: 10-slide Q3 marketing review for the leadership team. They know the strategy, they haven't seen the numbers. Goal: approve shifting 30% of the paid budget to content. Lead with the headline result, show what underperformed and why, end with the budget decision and what happens next quarter. One idea per slide, titles as takeaways, base it only on the data I pasted, no agenda or thank-you slides. Now the deck has a spine: the headline number opens, the failures get a slide because they justify the shift, the last slide is the decision — which is what the meeting is actually for. The bad prompt and the good prompt contain the same topic. Everything else in the good one came from the formula, and it took ninety seconds to write.

The follow-up prompts do the real work

Treat the first output as draft one, and budget two or three follow-ups. These work because they give the AI a decision to make instead of a topic to discuss — "make it better" changes nothing, while "merge these two" changes the deck:

Merge slides 4 and 5 — they make the same point.
Rewrite every title as the takeaway a listener should remember.
Cut the deck to 8 slides. Tell me what you removed and why.
Slide 3 is vague. Make the claim specific and list the data needed to back it.
Add a risks-and-mitigations slide right before the ask.
Which slide would a skeptic attack first? Strengthen it.
Move the recommendation to slide 1; make the rest supporting evidence.
Replace every generic example with a [NEED DATA] placeholder I can fill.
Turn slide 6 into a table: option / cost / risk / owner.
Write speaker notes: 60-90 seconds per slide, conversational,
each ending with a one-line transition to the next slide.

A pattern worth noticing: most follow-ups are moves — merge, cut, reorder, convert, strengthen — not rewrites. If you find yourself asking for "better" rather than naming a move, the problem is usually upstream in the formula, not in this draft.

A filled formula, end to end

Here's the formula filled for a quarterly review, so you can see what "two minutes of filling" actually produces:

[ROLE]     You are a presentation designer writing for a B2B SaaS company.
[AUDIENCE] The audience is the leadership team: 8 people, they set the
           budget, they know the strategy from last quarter's offsite.
           They care about whether marketing spend converts to pipeline.
[GOAL]     After the talk they should approve shifting 30% of paid spend
           to content, with the change live next month.
[INPUT]    Base it only on this material: [paste: metrics table, the
           paid/content breakdown, the two test campaigns, notes].
[FORMAT]   10 slides for a 15-minute slot. One idea per slide. Bullets
           under 12 words. Titles are takeaway sentences. Speaker notes
           on every slide. Results vs targets as a table on slide 2.
[TONE]     Direct, numbers first. Never use: synergy, leverage, robust,
           "in today's fast-paced world".
[AVOID]    No agenda or thank-you slides. No invented statistics — mark
           gaps as [NEED DATA]. No "we're excited to share".

Notice what the specific lines bought. "They set the budget" changed the deck's whole posture — this is a spending decision, so costs lead and the first slide is a cost curve. The banned-phrase list came from last quarter's deck, which a reviewer called consultant parody — the retro habit from the workflow article feeding the formula. The table instruction on slide 2 was spliced in from the QBR template. A prompt filled to this depth doesn't need luck; every failure mode in the troubleshooting table above is pre-closed before the first token is generated.

Prompt patterns that backfire

Some additions feel useful and reliably hurt. "Be creative" produces gimmicks — alliteration, forced metaphors, a "creative hook" slide that wastes the first minute; creativity in decks comes from the specificity pass, not from instructing flair. "Make it professional" produces the jargon inflation described in mistake eight of the mistakes list — the model's idea of professional is consulting-parody register. "Include best practices" is the generic-content machine: best practices are what the AI defaults to anyway, and requesting them converts your deck into a summary of blog posts. And a prompt twice the length of the formula with ten instructions degrades output the way ten goals do — if you're past the formula plus one template's structure, split the job into two prompts (structure first, then a rewrite pass) instead of one paragraph that asks for everything and prioritizes nothing.

Using the formula inside generators

Slide generators' prompt fields are shorter than a chat window, so the formula compresses without losing its job: the audience line, the goal line, the slide count, and the avoid-list are the four that must survive; role and tone can usually go. Paste your source material if the field allows it — Gamma and Copilot accept documents — and if it doesn't, generate the outline in your chat AI first and paste the outline instead, because the outline carries the input's value anyway. One generator-specific line worth adding: "use my outline exactly; do not add slides." Generators treat prompts as inspiration and will pad, which is the coverage problem wearing a nicer outfit. Then the generator does what it's good at, layout, while the formula keeps doing what it's good at, aim.

The formula for restructuring an existing deck

The same seven parts restructure a deck that already exists — yours from last quarter, or a colleague's draft you inherited. INPUT becomes the existing deck's full text, and the GOAL line does the heavy lifting, because restructuring is deciding what the deck is for. The prompt: "Restructure this deck so it argues for [goal] for [audience]. Return a slide-by-slide plan first: what to keep, what to merge, what to cut, what to add — before rewriting any slide text." Working from the plan is what makes restructuring safe: you approve the moves at the cheap stage, then the rewrite executes them. The rewrite prompt: "Now rewrite the slides per the plan; keep all verified numbers exactly as they are." This pairing has saved more drowning decks than fresh generation has — most work decks don't need a new deck, they need the existing one pointed at a decision, and the plan-then-rewrite structure keeps you in control of exactly which moves happen.

How the formula scales with deck size

A five-slide update and a forty-slide training use the same parts in different proportions. At five slides, the formula shrinks to three lines — audience, goal, format — because the input is short and the avoid-list has nothing to bite on; the deck is over before the defaults can appear. At forty slides, three parts expand: FORMAT gains section structure ("six modules of six to eight slides each; one agenda slide per module is allowed"), INPUT becomes a per-module paste — quality sag on very long inputs is a content-layer fact, not a tool one — and AVOID gains the training-specific bans: no walls of text, one concept per slide, jargon explained on first use. The parts that never shrink at any size are audience and goal: they're the aim, and aim doesn't scale down with deck length. If you only have thirty seconds to spend on a prompt, spend them on those two lines — everything else is configuration, those two are direction.

The full chain, end to end

The formula slots into the 30-minute workflow like this: the brief's audience and goal lines are the AUDIENCE and GOAL parts, already written — copy them in. Add INPUT from your source material, keep the FORMAT and AVOID lines as your standing defaults, and generate. Two or three follow-ups later, the draft goes to the human pass described in the editing article. Prompt work for a standard deck: about twenty minutes, most of it in the follow-ups, which is exactly where the quality gets made.

Save it as a snippet

Keep the formula in a notes file or a text expander. Filling it takes two minutes when a brief exists, and the second deck's formula is mostly the first deck's with new facts — the FORMAT and AVOID lines don't change between decks at all. Over a month you end up with a small library of filled prompts per deck type, which is the point where prompting stops being a skill you exercise and becomes infrastructure you reuse.

The formula doesn't make the AI more creative — it makes it aimed. The creativity, the specifics, and the judgment are yours; the seven lines just ensure the machine spends its effort on your problem instead of on the average of the internet.