Step 4 · Close the gap with the reference

Don't say "make it better". Make it compare the generated deck against the reference image, diagnose the gap itself, and then fix it directly.

The official tutorial puts it well: this self-comparison loop is often the moment the deck starts to feel designed rather than merely generated.

Why this works

"Make it better" has no reference object. The model has to guess which dimension you meant — colour? spacing? type? Guess wrong and you say it again.

Give it a reference and the problem becomes diagnosable: the type contrast here is 1.4×, the reference is 3×; every slide has the same whitespace, the reference alternates dense and open. Concrete gaps have concrete fixes.

This is also why step 1's image has to be kept. It isn't only for choosing a style. It's the ruler for the entire second half of the workflow.

The prompt

Here is the original reference image again.

Compare the current deck against it. Look at:
1. Layout structure
2. Type scale and personality
3. Colour palette
4. Spacing and alignment
5. Visual hierarchy
6. Texture, depth, and atmosphere
7. Whether every slide belongs to the same design system

Don't just tell me what's different. Improve the deck directly.

Use everything you have — design and code — to close the gap, while
keeping the deck readable and editable.

Then summarise:
1. What you changed.
2. Which slides improved most.
3. What still needs a human to review.

Why these seven

Between them they cover the whole distance between "designed" and "assembled":

Dimension	What weak looks like
Layout structure	Every slide is title-plus-body, no variation
Type scale and personality	Heading and body barely differ
Colour	The reference's colours are there, but the ratios are wrong
Spacing and alignment	Elements float free of any shared baseline
Visual hierarchy	Three things equally weighted; nobody knows where to look
Texture and depth	Entirely flat where the reference had grain and layering
Same design system	Slide 3 and slide 8 look like different authors

The last one fails most often. Generating many slides, an agent tends to optimise each locally, and the deck stops reading as one object.

"Don't just tell me what's different"

Keep that sentence. Without it, most agents return a diagnosis and stop, waiting for you to say "go ahead then" — a round trip that costs you time and credits for nothing.

"Readable and editable"

Keep that too. Told to use everything it has, an agent has two characteristic ways of overreaching: converting text to images for finer layout control, and stacking effects until the text stops being readable. That clause is the boundary.

What to check afterwards

In its summary, item 3 — what still needs human review — is the most valuable. It's usually where it wasn't sure either.

Then look yourself:

• [ ] Does it read closer to the reference's register?
• [ ] Was readability sacrificed for resemblance (type too small, contrast too low)?
• [ ] Did any slide get worse?
• [ ] Is the text still real text?
• [ ] Is there density variation now, or is every slide still the same?

A second pass is allowed

This step repeats. Make round two narrower:

Compare again. This round, two things only:

1. Type scale. The heading-to-body contrast in the reference is far
   stronger than what we have. Push it — body can go smaller, headings
   larger.
2. Density. Every slide currently carries the same amount. The
   reference alternates full slides with nearly empty ones. Redistribute
   by importance and give the key slides more air.

Change nothing else.

"Change nothing else" earns its place. Without it, round two may quietly re-lay-out what round one got right.

Where it goes wrong

Trap	How it shows	Fix
Diagnosis without action	You get a report	Keep "don't just tell me what's different"
Text becomes images	Nothing selectable	Add "keep all text as real text nodes"
Each pass gets busier	Decoration overtakes content	State that readability outranks visual impact
Too much in one pass	Some slides better, some worse	Split into rounds of one or two dimensions

Next

Step 5 · Real logos and brand assets — the visuals are there; credibility is next.