[🤖] After Days of Struggling with the Brochure, I Finally Found the Answer to AI-Made PPTs

✨ GPT-5.6 Sol’s Summary

No matter how much I gave it DESIGN.md and premium references, the PPT kept looking dated. Once I chose a finished image first, made it the visual source of truth, and had Codex turn it into an editable slide, the workflow that had blocked me for days finally opened up.

The Brochure Kept Coming Back as Garbage for Days

I spent days trying just about everything while making a hotel sales brochure for work. I found premium hotel brochure references and threw them in, made a DESIGN.md that organized the colors, type, and spacing, split out the content and instructions for each page, and even ran a pile of coordinator and worker sessions.

But the results kept looking dated and strange. Even when I had the same page rebuilt, it looked like an internal report or an old proposal instead of a premium hotel sales brochure. Codex would proudly return checks saying that no text overflowed, the file opened correctly, and the fonts matched. Then I would open the PPT and the first thing I felt was, “Ugh, what is this?”

When I decided I needed DESIGN.md because every brochure came out with a different design, I thought it would at least reduce how often AI had to guess the mood again from a blank canvas. DESIGN.md really was necessary. But this time I learned that simply having the file does not automatically produce good design.

Codex reported that it had read DESIGN.md and the references, but the actual PPT failed to capture why those references looked premium. It borrowed a few colors, made some similar boxes, and considered the references applied. The composition, area ratios, visual weight, spacing, eye flow, and overall impression disappeared. All that remained was a boxed document with neatly arranged content.

When the Code Came First, the Resulting Screen Became an Afterthought

As I kept watching the process, the design approach itself started to look wrong. It did not begin by imagining how the finished page should look. It broke the content into pieces, set coordinates for text boxes and rectangles, added lines and colors, and completed the PPT code. Only after that did it render the slide and reveal a garbage result.

Then it started changing the spacing, reducing colors, and moving boxes around. But it was still strange because those small edits preserved a composition that had been wrong from the start. The render was not a design target. It was closer to a debugging screen used to discover problems after the code was already written.

Codex is incredibly good at finding patterns, so it felt strange that no matter how many premium references I threw at it, it could not find the pattern of “premium” inside them. Thinking about it now, explaining the traits of good design and creating that complete impression from a blank PPT are different abilities. A result can pass every item for alignment, contrast, spacing, fonts, and color and still look painfully dated.

The Moment I Started with an Image, It Suddenly Designed Well

Then I told it to remake the same page as an image first using the color combination I had chosen. Absurdly enough, it produced a good image. The information hierarchy was visible, the balance of areas and spacing worked, and the icons and lines shared one coherent style. The modern, simple, premium hotel atmosphere I had kept asking for finally appeared.

“It makes a good image…? Then why does the PPT look like that?”

That was when I saw the answer. Instead of translating DESIGN.md and the references directly into PPT code, I should first have it create a finished page image. I can keep revising at the image stage until I like the result, then freeze that image as the page’s “visual source of truth.” The original PPT and page instructions remain a separate “content source of truth” that protects the wording, facts, relationships, and required items.

After that, I do not ask Codex to think up a new design. I ask it to reproduce the approved image as closely as possible with editable PPT elements. At the end, I render the saved PPT back into an image and compare it beside the visual source of truth at the same size.

The page rebuilt with this method was clearly different from the earlier results. It preserved the original wording and organizational relationships while bringing the approved image’s composition, color ratios, typography, icons, and lines into an editable PPT almost as they appeared. When I saw the result, I genuinely said, “Wow…! I found the answer!!!”

In the End, I Have to Review the Design Myself

After the day I failed to receive the brochure because I was too busy managing coordinators and workers, I became deeply suspicious of parallel work itself. Looking at it now, the coordinator-worker structure is not always bad. It works well for quickly generating image candidates in several directions or for turning separate approved pages into PPTs.

But I cannot leave design selection and final review to the coordinator. The coordinator can check missing copy, incorrect relationship lines, text overflow, fonts, and file integrity, then prepare comparison images. I have to decide for myself which result looks more premium, which page might make a hotel manager want to sign a contract, and whether I genuinely like the result.

From now on, I will make brochure pages in this order.

1. Freeze the content using the original PPT and page instructions.
2. Create several image concepts from DESIGN.md and the references.
3. Review them myself and approve the image I like as the visual source of truth.
4. Reproduce the approved image as closely as possible in an editable PPT.
5. Compare the PPT render and visual source of truth at the same size, then give final approval myself.

Before, I made the PPT first and discovered the design for the first time in the render. Now I finish the design in an image first and treat the PPT as the implementation that carries it over.

For days, I felt like I was going mad wondering why this would not work. At least now I know where the design has to be decided and who has to review it. The answer was not to make DESIGN.md longer or command it more forcefully to “work like a designer with 30 years of experience.” The answer is to create the finished image I choose with my own eyes first, then make Codex implement it precisely. For now, this is the best method I have found.