Don’t ask AI to make your presentation yet

Give AI the material, as well as your view.

Start with the files you would use yourself: the current spreadsheet, project notes, relevant research and perhaps the previous deck. Tell AI which sources contain the current facts and which are there only as examples of style or structure. Keep dates and definitions attached to the numbers.

Add a few sentences about who will see the presentation and what you need from them. Say whether you will present it, send it to be read, or both. Mention any required template and whether colleagues need to edit the result.

Then ask AI what the material actually supports. This gives you something to discuss before a plausible story hardens into twenty slides.

Help me prepare a presentation about [topic] for [audience].

What I need from them: [decision / understanding / discussion / action].

It will be [presented live / circulated to read / both].

I have attached:
- [source]: current facts
- [source]: background or context
- [source]: example of style or structure only

Treat [source] as the source of truth if materials disagree, but flag the disagreement rather than silently resolving it.

Preserve dates, definitions, units and reporting periods attached to numbers.

The eventual output needs to be [PowerPoint / Google Slides / HTML / PDF / other], using [template or reference if any], and [must / need not] remain editable by colleagues.

Before making slides, tell me what the material supports, what conflicts, what appears to be missing, and what you would want to clarify. Cite the source and location for important findings. Do not invent missing facts.

You do not need a settled opinion to begin. A question and a few competing explanations are enough.

Suppose you are briefing leaders on whether to expand a six-week pilot of a new workflow tool. You currently think the pilot is promising, but that use varies too much between teams and the benefits have not been measured well enough to justify expanding it yet.

Do not ask AI to turn that view into slides. Ask it to test it.

Use the attached usage report and pilot notes.

I’m briefing leaders on whether to expand the pilot. My current view is that use is uneven and we have not measured the benefits well enough to justify expansion yet.

Check that view against the material, including evidence against it. What alternative interpretation would a reasonable reader make? What evidence would change the recommendation?

Do not make slides yet.

The point is to let the evidence change your thinking rather than ask AI to decorate it.

Ask for a slide plan you can check.

Once the story is taking shape, ask for a short deck plan: what is each slide here to do, what should people take away, what supports it, and how might you show it? This is where an outline becomes specific enough to build.

Part of the pilot example might look like this:

These are proposed takeaways to check against the sources, not findings to assume. Look for slides with no evidence, gaps in the story, and several slides doing the same job. Fix those in the plan.

Using the attached sources and this working view: [paste], propose a deck plan for [audience and purpose].

For each slide, give:
- its job
- the takeaway or question
- supporting evidence, with source location
- the likely visual

Mark missing evidence.

Suggest another sequence only if it would materially improve the story. Do not design slides yet.

Make sure the story works before the layouts do.

A title-only pass is a cheap way to expose jumps, repetition and buried decisions. It tests the sequence, not whether the claims are true.

Here are my slide titles in order: [paste].

The audience is [who] and the purpose is [what].

Tell me what story they convey and where it jumps, repeats or leaves a question unanswered. Suggest only necessary changes.

Do not infer that a claim is true from its title.

Try the difficult slides before building the rest.

For a new deck, I would prototype three representative slides: the opening or recommendation, the hardest numbers-heavy slide, and a slide explaining a process or idea. Use real content, including the awkward labels and missing data. A beautiful cover proves very little.

Start with what the audience needs to see. If the point is uneven use, show the differences between teams; a large total-user number will hide them. If you are comparing options, use the same criteria for each. If a process is slow, show where the time goes.

This slide needs to show [point].

The supporting material is [paste or attach].

Suggest two ways to display it. Explain what comparison or relationship each makes easy to see, and what it might hide.

Preserve units, time periods and uncertainty.

Recommend one and explain why. Do not add decorative imagery.

Then settle the styling. Start with your firm’s template or a few examples you like. Tell AI what to borrow: readable charts, clear labels, restrained colour, the treatment of photographs. Ask it to preserve those choices across the deck without forcing every slide into the same layout.

Taste and Impeccable are useful references for making design choices explicit and reviewing the result. Borrow that approach rather than importing website-design rules wholesale.

Use the attached deck plan, source material and template or visual references.

Build three representative slides:
1. the opening
2. the hardest data slide
3. a process or explanation slide

Use this prototype format: [PowerPoint / HTML / other].

Keep sources and unresolved gaps visible.

Export images and inspect them if you can. Otherwise say what remains untested.

Stop before building the full deck.

The useful habit is testing a small sample, not using a particular file type. If HTML is the final presentation, design the PDF deliberately too. If you need PowerPoint, preserve editable text, charts and tables rather than delivering a pile of screenshots.

Once the plan and sample work, let AI produce the rest.

It should not need permission for every text box. This is the point where delegation becomes useful because the consequential choices have already been surfaced.

But the plan is not a contract. If a chart reveals a problem with the argument, revisit the story rather than preserve a bad slide because it was approved earlier.

Using the approved deck plan, sources and the three prototype slides, build the remaining presentation.

Preserve the argument, visual system and information-design choices established in the prototype, but do not force every slide into the same layout.

Every important claim and number must remain traceable to its source. Mark unresolved gaps rather than filling them.

Render the complete deck and inspect it as a whole and slide by slide. Fix obvious layout problems, repetition and inconsistencies.

If the evidence forces a change to the agreed story, stop and explain the issue rather than designing around it.

Find out what the deck says without you.

Export the slides as images or a PDF and inspect them. A whole-deck view helps you spot repetition; full-size pages let you check labels, wrapping and whether the important comparison is visible.

For a useful extra test, give the finished slides to a fresh AI conversation that has not seen the drafting discussion. Supply the audience’s role, but leave out your intended conclusion and deck plan. Ask it to reconstruct the argument before suggesting improvements.

Use only the attached slide images.

The intended readers are [roles].

Tell me:
- the main point
- the reasons supporting it
- what the readers are being asked to do

Identify slides where you had to guess or where the visual suggests something different from the title.

Do not rewrite anything yet.

Compare that account with what you meant. Inspect the slides it misread before accepting its diagnosis. You are testing what the deck communicates, not taking another model’s opinion as a verdict.

Check the facts separately

Do not combine factual verification with another round of copy editing. Go back to the original material. Ask AI to trace important figures and claims to their sources, reproduce calculations where possible, and flag discrepancies. Inspect the supporting material yourself for the claims the decision depends on.

Check the attached finished presentation against the original source material.

For every important number, factual claim and recommendation, identify the supporting source and location.

Reproduce calculations where possible.

Flag anything that is:
- unsupported
- inconsistent with the source
- based on an outdated period
- missing an important qualification
- presented more strongly than the evidence warrants

Do not improve the writing or design during this pass. This is a factual check.

A finished deck is not a rehearsed presentation.

If you will speak live, practise aloud with a timer. Use AI to test the questions you may face, including the awkward ones the slides avoid.

Use the attached deck and my speaking notes.

I have [minutes], including questions, with [audience].

Help me cut the talk to fit.

Then ask likely audience questions one at a time, including questions that challenge the recommendation.

After each answer, identify anything unclear or unsupported before moving on.

Finally, open the file or link as the recipient will. Check that it works, that the reading copy stands alone, and that nothing important depends on a hidden note or an untested interaction.

Save what you want AI to do again.

You should not need to explain your chart conventions and review standards every time. Keep a small reusable set of instructions: this workflow, firm requirements, a few good examples, useful slide patterns and mistakes worth preventing.

If your AI environment supports reusable Skills, turn those instructions into one. ChatGPT and Claude both support Skills, but availability, creation and installation differ by product, plan and workspace. Have the agent check what is available where you are working.

Start with a couple of decks, not an archive. Let AI suggest the patterns it sees, then approve what to keep. Your old slides may contain habits you would rather lose. Keep presentation-specific figures and conclusions out of the reusable instructions.

Read the article supplied with this prompt and any examples I attach.

Create a concise Presentation Skill covering the workflow, slide checks, firm rules, useful examples and the review process.

Ask one useful question at a time for genuinely missing information. An article-based draft can come first.

Get my approval before saving inferred preferences. Keep deck-specific facts out.

Verify what you can create or install here, and explain any remaining setup. If installation is unavailable, return reusable instructions and files instead.

Test it on a small task using material I provide. Report what was actually tested.

After a real presentation, save a few lessons from the corrections and questions you received. Those are more useful than asking AI whether it liked its own work.

Compare the attached Presentation Skill with the finished deck and this feedback: [paste corrections or audience questions].

Propose up to three specific changes that would help next time.

Show the exact instruction or example to change and why.

Do not update the Skill until I approve.

For a routine update, reuse the plan and layouts, refresh the sources, and review what changed. You do not need to restart the whole process.