Step 6 · Images that carry an argument

With the real assets in, the deck may still want conceptual imagery. These aren't brand assets. They're pictures that help explain a point, set an atmosphere, or make something abstract concrete.

The only hard part of this step is restraint.

One test

An image should explain the argument, not decorate the slide. If it doesn't make the idea easier to understand, remove it.

Use it as the acceptance criterion. Once the images are placed, ask of each: take this away — is the slide less convincing? If not, it's decoration.

The prompt

Review the whole deck and decide where an AI-generated image would make
the presentation clearer or more persuasive.

Instructions:
1. Do not add images to every slide.
2. Only add one where it helps explain an idea, supports an argument,
   or makes a key moment more memorable.
3. Skip slides that already carry enough visual content.
4. Keep the image style consistent with the chosen deck style.
5. If you can generate images, generate them and place them.
6. If you can't, write the image prompts I should take elsewhere.

For each image you add or recommend, tell me:
1. The slide number
2. What the image is doing
3. The prompt used
4. Why this slide benefits from an image

Those four questions are the point

Requiring a purpose and a justification per image forces it to apply the test itself.

In practice you'll see some justifications strain — "makes the slide more engaging", "adds visual interest". Those are the ones to cut. The specific ones stay: "this slide describes data flowing between three systems, and in text the reader has to assemble that structure in their head".

Slides that genuinely want an image

Situation	Why it helps
An abstract concept	Gives it a shape to point at
A process or a system	Spatial relationships beat prose
A comparison	Side by side beats two paragraphs
An emotional turn	"This is where we were standing"
Too many text slides in a row	A pacing problem — but pick the right place

Slides that shouldn't get one

• Slides that already have a chart (the chart is the visual)
• Slides that already have a real screenshot or logo
• Slides that are already dense
• Pure transitions (section breaks have their own treatment)
• The conclusion, which should be where the words carry the weight

Style consistency

Clause 4 isn't a throwaway. Drift is the most visible failure of AI imagery: flat illustration on slide 3, photoreal on slide 7, a 3D render on slide 11 — three decks stapled together.

Tighten it:

All images use one style: [describe it — e.g. "low-saturation,
monochromatic, grainy abstract illustration; no faces, no text"].

Generate one first. I'll confirm it before you make the rest.

Confirming one before the batch saves a round trip.

If your agent can't generate images

Clause 6 makes it write the prompts instead. Take those elsewhere and place the results by hand.

OpenDesign has image generation built in (which models depends on the release — GPT Image, Seedream, and Nano Banana are the usual ones), so clause 5 will normally apply.

Where it goes wrong

Trap	How it shows	Fix
An image on every slide	The deck becomes a picture book	Keep the clause, and set a maximum
Style drift	Every image looks differently sourced	Approve one, then batch
Text inside images	Generated lettering is usually garbled	Say "no text of any kind inside the image"
Images upstaging	The audience looks at art, not the argument	Shrink them, or move them to the margin
Images unrelated to content	Pure atmosphere	Apply the test and delete

A number to aim at

In a 10-to-15-slide deck, two to four slides genuinely want a generated image. If more than half of them have one, you've overshot.

Next

Step 7 · Charts that say something — on data slides, charts matter far more than imagery.