I Made This Deck With Codex

I had a keynote coming up, and I wanted to see how far I could push Codex.

So I gave it my messy source material, set up a slide-by-slide loop, and let it run.

About six hours later, I came back to this:

This is the part that made me sit up. It was not one lucky slide. It was a coherent section of a real keynote deck: abstract AI concepts, visual metaphors, diagrams, charts, and generated backgrounds that mostly felt like they belonged together.

It was also a native PowerPoint file. Not a PDF, not screenshots glued together, not a fake deck. The text was editable. The slides opened in PowerPoint. I could keep working with it like a normal office artifact.

That is what this post is about. I am going to show you the source material I gave Codex, a cleaned-up version of the loop I used, where it failed, and how to set up the same workflow yourself.

Brennan McEachran 👨‍🚀

@i_am_brennan

·Follow

Trying some ralph loops on gpt-5.5 w/ POWERPOINT. Image Gen skill + presentation skill. Slide by slide, loop: generate images, select, regen/edit it properly for a slide, integrate into pptx, review and verify correctness. Move to next slide. 🤞  Show more

1:55 AM · Apr 24, 2026

0

Reply

Copy link

Read 1 reply

The Short Version

If you ask a chatbot to "make me a deck," you will probably get something generic.

If you give an agent a workspace, source material, tools, a review surface, and a loop that keeps going slide by slide, you can get something much closer to a real deliverable.

The tool setup was not complicated. I ran GPT-5.5 in the Codex Mac app with the preinstalled Image Gen and Presentations skills. For this run, Image Gen used GPT-image-2 for visuals. Presentations let Codex write and edit an actual .pptx.

The real setup was the folder I handed it.

What I Put In The Workspace

I did not start with an empty prompt. I gave Codex messy, real materials:

• two Granola call transcripts
• a rough Google Slides outline with placeholder images
• a Wispr Flow talk track where I talked through the idea behind each slide
• the existing deck file
• a plain-text workpad the model could update as it went

The useful input was not polished. A simplified version looked like the stuff you normally leave scattered across docs, voice memos, and half-finished slides:

1Slide: related things land closer together
2
3Need to explain embeddings without making it feel like math class.
4Use examples people in the room will get: LeBron, lacrosse, Team Velocity, DMSC.
5The point is that meaning has shape. Related concepts cluster.
6Make it feel like a map, not a spreadsheet.

That is the trick: do not ask the model to invent what you have not given it. Bring the transcripts, voice notes, rough outlines, old slides, placeholder images, and rambling talk tracks. The model is far better at synthesis than invention.

Brennan McEachran 👨‍🚀

@i_am_brennan

·Follow

Core part of the prompt here. The rest is: a) 2 call transcripts from @meetgranola b) 1 google slides dot-dash deck with random images c) @WisprFlow'd talk track of the idea of each slide d) this explicit loop x.com/i_am_brennan/s…

Brennan McEachran 👨‍🚀

@i_am_brennan

Core part of the prompt. IMO gpt-5.4 doesn't really need a "ralph loop" if you are super clear about validation criteria. 5.5 seems the same

4:05 AM · Apr 24, 2026

0

Reply

Copy link

Read 1 reply

The Loop

This is a cleaned-up version of the prompt shape that made the difference.

I did not ask Codex to make the whole deck at once. I told it to run this loop one slide at a time:

10. Read the running notes file in full before starting.
2   Know what has been decided, what is in progress, what has been fixed.
3
41. Look at the slide. Understand what it is trying to say.
5
62. Generate a background image — high quality, conceptually accurate,
7   stylistically consistent with the rest of the deck.
8   No text baked into the image.
9
103. Place all editable text and labels on top in the actual PowerPoint slide.
11
124. Render the slide. Take a screenshot. Look at it.
13   Does it work — visually, factually, emotionally?
14   If not, go back to step 2.
15   If yes, go to step 5.
16
175. Update the running notes. Move to the next slide. Repeat.

Here is why each part matters.

The workpad is memory. I literally gave Codex a running notes file and told it to read it before each slide. That file became the whiteboard for the run. It tracked what visual language had been chosen, which slides had already been fixed, which images needed regeneration, and what criteria mattered.

The starter version can be almost embarrassingly simple:

1# Workpad
2
3Goal: finish the deck one slide at a time.
4
5Visual direction:
6- high-end keynote, not corporate template
7- generated visuals can carry the slide
8- all real text stays editable in PowerPoint
9
10Current state:
11- no slides completed yet
12- update this after every slide

The image rule keeps the deck editable. The model was allowed to generate beautiful backgrounds, diagrams, and atmosphere. It was not allowed to bake real slide copy into the image. All headlines, labels, and bullets had to live in PowerPoint as editable text.

That one rule is the difference between a working deck and a pretty poster.

The screenshot is the review surface. After each edit, Codex rendered the slide and looked at the screenshot. If the image was wrong, the layout was cramped, or the concept did not land, it had to repair the slide before moving on.

This is the part most people skip. They prompt, accept the first output, and call the model bad. The loop forces the model to look at its own work.

The loop is allowed to take time. This run took about six hours. That sounds slow until you remember I was not sitting there manually placing every object. The model was working through the deck while I could do other things.

Where The Loop Fixed The Deck

The clearest example was the diffusion slide. The starting point was accurate, but it looked imported. It had the feel of an academic diagram pasted into a keynote.

Before: accurate, but visually outside the deck.

After: the same concept, rebuilt in the deck's visual language.

The loop looked at the slide, decided the visual language did not match, generated a replacement image, inserted it into the native PowerPoint slide, preserved the editable labels, rendered the result, and checked it again.

It also caught problems. At one point, the verification log shows the model stopping to rebuild and render the slide before accepting the new asset:

This is what makes the loop useful: the model is not just generating, it is rebuilding and rendering before final verification.

That is the mechanic underneath the magic. The model is not just producing slides. It is producing, rendering, inspecting, and repairing an artifact.

A Few Slides

Quick Start Guide

Here is the practical version.

Step 1: Gather your trash. Make a folder with everything you have: transcripts, notes, rough outline, old deck, placeholder images, voice memos, and a talk track. Do not clean it up too much. The mess is useful.

Step 2: Write a three-sentence visual brief. Tell the model what the deck should feel like. For example:

1This should feel like a premium keynote, not a consulting template.
2Use generated visuals as the main emotional and conceptual layer.
3Keep all real text editable in PowerPoint; no baked-in headlines or labels.

Step 3: Create the workpad. Add an empty workpad.md or notes.txt file. Tell the model to read it before every slide and update it after every slide. This is how you stop a long run from drifting.

Step 4: Start the Ralph loop. Paste the loop prompt and make the instruction explicit:

1Work one slide at a time.
2
3Before each slide, read the workpad in full.
4
5For each slide:
61. Inspect the current slide and understand what it is trying to say.
72. Generate a high-quality background image with no text baked in.
83. Place all real text and labels as editable PowerPoint elements.
94. Render the slide and inspect the screenshot.
105. If it fails visually, factually, or emotionally, fix it before moving on.
116. Update the workpad with decisions, problems, and what changed.
127. Move to the next slide.
13
14Do not skip ahead.
15Do this until the full deck is complete.

Step 5: Let it run longer than feels normal. This is not a one-shot prompt. It is closer to assigning a junior designer a folder, a brief, a checklist, and a review cycle. Check in, redirect when needed, but let the loop do real work.

The transferable skill is not "make PowerPoint with AI." It is learning how to turn knowledge work into a loop a model can inhabit: context, artifact, tools, review, repair, repeat.