I'd Have Budgeted Three Days.
With AI, I Finished in Three Hours.

Around 3 p.m., I received an urgent request. The client needed presentation material for a meeting that day, but could initially send only a Korean-language text document. Images and supporting material would arrive as I worked.

I work in branding and design. For a brief like this, I would normally allow several days to read the material, arrange the story, design the pages, and revise them after feedback. Using my previous workflow, I would have expected to need at least three days.

This time, I finished that day with AI. About three hours later, the final corrections were in place, and I had both an editable PowerPoint and an HTML version to view in a browser. The document had grown to 15 pages.

That does not mean one request to "make a PowerPoint from this" produced the finished job. I changed the design after seeing the first result, refocused the document after feedback, and incorporated material and images that arrived later.

What surprised me was not the absence of revisions. It was reaching a finished version in about three hours despite all those revisions. This article follows the sequence of that afternoon.

Actual work examples have been edited to remove addresses, identifying location details, and client identifiers. Reconstructed diagrams are labeled as illustrative examples.

TOOL GUIDE · Select a name to see its role

ClaudeContent into a first draftCodexCreate and revise filesImage generationSet the cover atmosphereA. (Adot)Capture spoken feedbackAsideBrowser connection and review

What I delegated in this project

Content into a first draft

+ Client manuscript+ Existing material

The key request

Organize the story for a client meeting and turn the content into a PowerPoint.

What came back8-page PPT · HTML

A first version with the content organized into pages

What I checked

I reviewed alignment and photo proportions to decide which parts needed redesigning.

Requests and results are generalized to show the workflow. Original client wording and confidential information are not included.

I first gave the material to Claude to organize the document and produce HTML and an eight-page PowerPoint. I started by handing over the work of moving material onto blank pages.

The first result was not ready to use. Photo proportions and alignment needed attention, and although the information was there, it was not clear where to look first. After one revision, I passed the PowerPoint to Codex to work on the design.

There were also moments when I had to check file access while handing the work between tools. I checked the Aside browser connection at that point, then continued the redesign by reading the original PowerPoint directly. Even connecting the tools did not happen flawlessly in one step.

My job was to look at the draft and ask, "Could I use this material to explain the subject to a client?" Once AI had put the content onto pages, I could respond to something in front of me instead of starting from a blank canvas.

The part I considered most important was choosing the design reference. Even when AI creates the document, the composition and design you show it make a substantial difference to the result.

"Make it clean" does not adequately explain what I want. A document built around large images needs a different layout from one designed for quick comparisons of numbers and tables. I looked beyond whether a reference looked attractive and asked whether its design supported the explanation this document needed to give.

For this job, I gave Codex images from a proposal with the feel I wanted. I pointed to clearly defined headings, generously sized key images, and tables with consistent spacing. I also supplied the established colors and typeface.

I did not stop at "make it feel like this." I specified how to distinguish headings, how much space images should occupy, and the order in which the tables should read. Rather than leaving every style decision to AI, I asked it to work within a direction I had chosen.

Borrowing another document's content is different from using its design as a reference. I was not transferring company names or copy. I selected design criteria relevant to this material: the contrast between heading and body sizes, the proportion of the page given to images, the spacing, and the alignment.

I generated the cover background with an image-generation tool. I left space for the title and added the text and logo separately so they could be edited later. For images needed to explain facts, however, I used material supplied by the client. I distinguished imagery that creates atmosphere from material that serves as evidence.

01 / The generated background

02 / The actual cover layout

With those criteria, I received a redesigned ten-page PowerPoint and HTML document. I reviewed the alignment and image proportions that had disappointed me in the first draft, then kept subsequent revisions within the same direction. The reference was not only a starting point. It was also a standard to return to during revisions.

During a call with the client, I clarified what the document needed to explain first. Including every piece of information was less important than putting the decisions the reader needed to make in this meeting at the front.

We were using the same material, but changing the order of explanation. I checked the call summary and transcript in A. (Adot), then passed the relevant revisions to Codex. This reduced the effort of trying to reconstruct feedback from memory after the call.

I did not simply forward the summary without checking it. If confirmed facts and plans still under consideration become mixed together, a reader can take away the wrong meaning. I checked the surrounding conversation and reviewed whether the client's priorities were reflected in the explanation order.

AI applied the revisions to the document. I checked what needed to change and how.

Maps and site photographs arrived while I was working. Showing three parcels separately did not make their relative locations clear, or show which explanation belonged to which parcel.

At first, I placed the material side by side. That still made the reader switch between images to match the locations. So I asked for all three parcels on one map, each in a different color, with a matching line connecting it to the explanation on the right. I used the same colors on the location map and the cadastral map.

MAP ANNOTATION / Explain with color and lines

Location mapParcel layout

Select an image to enlarge it.

AI-generated fictional maps illustrating the annotation method. Addresses, geography, and parcels are unrelated to the real site. These are not cadastral or survey records.

The illustrations above use fictional addresses and geography to demonstrate that method without disclosing the actual site. I also grouped site photographs that needed to be viewed together on one page. Rather than just adding pages, I focused on reducing the work readers had to do to interpret the connections between the materials.

At this stage, my request was not simply "make the images bigger." It was "make the relationship between these materials clear at a glance." Once I had clarified what the page needed to show, we could adjust the layout accordingly.

As more content arrived, the document grew to 15 pages. Instead of shrinking text to squeeze everything onto a single page, I split longer explanations across pages. I reduced overly tall table-header rows, removed paragraphs that were unnecessary for this meeting, and revised page titles at the end.

I explained what needed changing and why, and Codex applied those changes within the existing design. I did not need to transfer all the content again or rebuild every layout. I kept reviewing each result and deciding what to change next.

About three hours after starting, the final revisions were complete. I prepared an editable PPTX and an HTML document that could be paged through in a browser.

Looking back, few stages were finished in a single pass. I revised the first design, changed the content after the call, incorporated additional material, and checked tables and titles again.

What changed was that I no longer had to repeat every production step myself. I shared the work of transferring content, changing layouts, and producing revised files with AI. I could focus more on choosing references, understanding the client, and checking whether the reader would understand.

The reference was the starting point. I first chose a composition that could explain the content well, then told AI which design criteria to take from it. For me, a reference was both a direction to give AI and a standard for judging its output. I also obtained the necessary supporting material from the client. This was not a project in which AI found and decided everything from scratch.

I finished in about three hours a job I would previously have expected to take at least three days. This was not a controlled comparison of the same task using two methods. But because it was a real commission that I took through to a final version, the difference felt significant to me.

It was not fast because AI got everything right at once.
What changed was being able to make the necessary revisions
and still finish that day.

Explaining what to delegate mattered as much as choosing a tool. These are the four things I want to clarify at the start of my next project too.

1. This document is for [reader] to use for [purpose].
2. The key message is [message]. Confirmed material: [files]. Material to come: [material].
3. Use the reference for [heading hierarchy, image proportions, table spacing], while keeping the original content.
4. Use [typeface and brand colors] and produce [file formats]. Check image proportions and text clipping.

If some material is still missing, distinguish what is confirmed from what will arrive later. When sharing a reference, go beyond "make it like this" and specify the useful elements, such as the contrast in heading sizes or the image layout. For revisions, describe the difficulty: "the material and its explanation are too far apart to compare" is more useful than "make it cleaner." That explanation helps both AI and a human designer.

I have seen many claims that AI makes work faster. Taking a real client brief through to completion, while requirements change and new material arrives, is another matter.

This time, I tried it myself. The first output was not exactly what I wanted, but after several revisions I reached a point I could call finished. What stayed with me most was completing that day something I would previously have scheduled several days for.

I will need more experience before knowing which other projects can move this quickly. But the next time an urgent request arrives, I do not think I will judge its feasibility only by how long my old workflow took.

Have you used AI to finish a real project? I would love to know which stage made the biggest difference for you.