Turn a Report Into a Slide Deck With Three AI Models

Why does an AI slide deck usually come out wrong?

Because the deck gets requested before the argument is agreed. Hand a model a 20 page report and the instruction "make a presentation" and it will do the thing it is genuinely good at: compression. It walks the document top to bottom, shrinks each section into bullets, and returns a deck with the same shape as the report. That is a summary with slide breaks in it.

A presentation has a different job. It has to move one specific room to one specific decision inside a fixed number of minutes. The report's structure is almost never the right order for that, because a report is written to be complete and a deck is built to be persuasive. So the first question is not "what goes on slide four". It is "what does this room have to walk out believing, and what is the shortest honest path to it".

Should you ask for the outline or the slides first?

The outline, always, and one line per slide. Ask for a slide list where each entry states the claim that slide makes, not the title it carries. "Q3 results" is a title. "Growth came from renewals, not new logos" is a claim. A deck assembled from claims has an argument running through it. A deck assembled from titles is a table of contents with a background image.

Give the models a short brief before they touch the report: who is in the room, what decision you are asking them to make, how many minutes you have, and what they already know. Write that brief in Solo first, with a single model, because debating the brief across three models spends the disagreement you want to save for the storyline round. Anthropic's prompting guidance is direct about this, that stating the task, the audience and the output format explicitly beats implying them (be clear and direct). A deck brief is exactly that instruction, written once and reused by every model in the chat.

How do you run the outline round across three models?

Open aiDex, attach the report in Documents so every model in the conversation reads the same source, and switch to Compare. Compare sends one prompt to several models at once and lays the answers side by side, which is what you want here: three independent readings, not one reading plus two agreements.

Pick models from different lineages in the Dex, for example Claude Opus 4.8, GPT-5.4 and Gemini 3.1 Pro. Three variants of the same family tend to reach for the same headline, so the disagreement you are paying for never shows up.

Then send one prompt:

Here is the report and the brief. Propose a 10 slide storyline. One line per slide, and each line must state the claim that slide makes. At the end, mark the single slide you would keep if you had two minutes instead of twenty.

That last sentence does most of the work. Read the three answers by looking first at where each model put its mark. When all three keep the same slide, you have found the spine of the deck. When they keep three different slides, your report is making three arguments at once and the room will feel it.

How do you pick the winning storyline without arguing about taste?

Write the criteria before you read the outlines, otherwise you will pick the one that sounds most like you. Five criteria cover most decks:

Now switch to Judge. Paste the three outlines and the criteria, and ask for a score per criterion with one sentence of reasoning each. Judge is useful here not because the number is authoritative, but because the reasoning tells you which criterion each outline failed on. Very often the best deck is outline two's spine with outline one's opening, and you only see that once the scoring separates structure from sequencing.

How do you turn the winning outline into real slides?

Use Pipeline, which passes the work through a chain of models in fixed roles: Draft, Critique, Revise, Polish. For a deck, give each stage a narrow instruction.

Draft expands the approved outline into slide text with a hard limit, for instance a headline of at most 12 words plus three supporting lines. Critique reads the draft against one question only: does each slide prove the claim in its headline, or does it merely restate it? Revise fixes the slides Critique named. Polish enforces the format rules, parallel phrasing, consistent tense, numbers formatted the same way everywhere.

The limit matters more than the wording of the prompt. Without a cap, every model writes paragraphs and calls them bullets. With a cap, the model has to decide what the slide is for, which is the decision you actually wanted it to make.

How do you get speaker notes that are not the slide read aloud?

Switch to Team, where the models talk to each other in one conversation with a moderator running the floor. Ask for two things per slide: the sentence you say when the slide appears, and the question this slide invites from the room.

Then give one seat a job nobody wants. Tell one model to act as the most skeptical person in the meeting and to attack the deck slide by slide. The questions it raises become your Q and A preparation, and the slides that cannot survive the attack usually need a source, not a rewrite. You can keep a working panel like this in Teams and reuse it for every deck you build.

What should the panel never decide?

Three things. It should not invent or adjust numbers, because a figure that is not traceable to the report is a figure you cannot defend in the room. It should not decide what is confidential, since no model knows your disclosure rules. And it should not choose the ask, because the thing you are requesting from that audience is a judgment about relationships and timing, not a writing task.

Everything else, structure, sequencing, phrasing, compression, is exactly where a panel beats a single chat.

What if the report is confidential?

Run the whole routine locally. aiDex connects to Ollama models running on your own machine, so board material, salary data, unannounced results and client documents never leave it. You lose some of the range you get from frontier models, but the routine survives, because the value here comes from independent readings and a scoring pass, not from any one model's brilliance.

For everything else, the panel is yours to configure. Use your own provider keys or the ones we manage, and pick the models you want.

Where should you start?

Take the last report you presented badly. Write the four line brief, attach the document, run one Compare round for storylines, and compare the marked slide across the three answers. If they disagree, you just found out why the presentation did not land, and it was never the slide design.