Making Slide Decks with AI — How to Build a System So You Stop Redoing Every Deck

For a long time, there were two ways to make course materials: build them yourself, one slide at a time, or hand the job to an AI slide generator.

I’ve tried both. Making slides myself cost me time, and handing the job to a generation tool (I used Genspark) meant ongoing costs, and results I had to fix again every time. My standard template uses a 4:3 ratio with set colors and a header structure, and the slides that came out were never that template. The third problem was more exhausting: I had to repeat the same revision instructions from scratch every time. What I said yesterday, I had to say again today.

So I changed my approach. Instead of asking AI to make slides, I turned my production method into a document and handed it to AI. In one day I got 16 slides for a special lecture and 70 slides for a 4-hour course, and, more importantly, a system I can keep using.

Below is exactly the order I followed. I wrote no code at all, and I used just one tool: Claude Cowork (an AI workspace that handles files and folders). The principle is the same if you use another AI tool.

Step 1 · Start with the folders (this is half the work)

Don’t start by agonizing over prompts. The folder structure becomes your workflow. The AI looks at this structure to work out which stage you are at. Here is the structure I use.

CourseMaterials/
├─ 00_Guidelines.md          ← where you write down your rules (the key file)
├─ 01_Template/              ← one original standard template PPTX
├─ 02_Samples/               ← finished PDFs I edited myself
├─ 03_IllustrationLibrary/   ← reusable images
├─ 10_ByCourse/
│   └─ 2026-08_○○Training/
│       ├─ 1.Planning  ├─ 2.Draft  ├─ 3.Illustrations  └─ 4.Final
└─ 99_Archive/

• 01_Template is the original you never touch. New materials start by duplicating it here.
• 02_Samples plays a big role later. Put finished work that I polished myself in here, and the AI learns my taste from it.
• 03_IllustrationLibrary is an asset that grows in value with every session. It holds 18 items now, reused in new courses.
• 1. Planning → 2. Draft → 3. Illustrations → 4. Final — these four slots are the production stages. Which slot a file sits in tells you its progress.

Once the folders are ready, connect only this one folder to your AI workspace. Don’t connect your home directory or your entire cloud drive — I explain why separately below.

Step 2 · Have AI break down your template

The first instruction isn’t “make it.” It’s “read it and organize it.” I had AI open the template file I’ve used for over 20 years and extract its specifications.

What came back was a table: slides are 4:3, my navy color code, the fonts, and the role each of the 16 pages plays (cover, contents, body, summary, and so on). It was the first time I saw what I had been doing by instinct in actual numbers.

Then I added one more line of rules. That one line ended the “fixing it again” cycle.

Every slide deck after that matched my format exactly. In AI terms, this is a system prompt or context engineering, but all I built was a single human-readable document.

Step 3 · What to put in the guidelines file

The guidelines file (00_Guidelines.md) is the heart of this system. The format is up to you, but I divide mine into six sections.

1. Absolute rules — Two or three lines you can never break (duplicate the template, no blank screens, no unverified figures).
2. Template specs — the measurements the AI extracted in Step 2.
3. Production process — The folder flow (Planning → Draft → Illustrations → Final) and what to do at each stage.
4. Command dictionary — Recurring tasks bundled into a single phrase. (Explained in Step 4)
5. Style rules — Where my revision instructions accumulate. (Step 5 is the key)
6. Change log — a table of dates and changes. Keep it so you can trace later why each rule exists.

Don’t try to finish it all at once. Fill in just items 1 and 2 and start; the rest builds up on its own as you work.

Step 4 · Turn recurring tasks into “commands”

If you’ve made the same request three times, it’s time to turn it into a command. Define it like this in the command dictionary in your guidelines.

### Command Dictionary

**“Archive it”**
1. Check the images used in this lecture in 03_Illustration Library
2. Report the list of files no longer needed first (do not delete right away)
3. Move only the approved files to 99_Archive
4. Tell me if anything should be reflected in the guidelines

**“Insert illustrations”**
Insert the images with the specified file names from the 3. Illustration folder into the exact positions on the slides

Now a single phrase, “archive it,” runs all four steps on its own. Explanations don’t just get shorter — they’re no longer needed at all. You’ll notice I put “don’t delete right away” in step 2 — for any action that deletes, it’s safer to build in a reporting step first.

Step 5 · Don’t let feedback evaporate (this is the key)

Once a draft comes out, you give revision instructions. “No illustration on the cover.” “Same top and bottom margins.” “Bigger text.” “Center table contents vertically.” “Make the keywords much bigger.” “Keep explanations to one line.” Normally, you fix these and move on. That’s why you end up saying them again next time.

Add just one sentence to your guidelines and this loop ends.

That one day alone added 9 rules. And I never had to make the same correction twice. The more sessions you run, the fewer instructions you need.

One step further: add finished work and let the AI learn from it

I put a finished PDF I had edited myself into 02_Samples and gave this instruction.

After comparing 71 pages, the AI pulled out 5 rules on its own, including how to place illustrations (large, at the very end, bottom-aligned), the color system for underline emphasis, and how to handle screen captures. I had never explained any of them. I recommend keeping your finished decks together in one folder rather than discarding them.

Step 6 · Divide the work between people and AI

• Illustrations — the AI leaves numbered placeholders in the slides and writes image-generation prompts for you. I generate the images with an image tool (I use Nano Banana) and just drop them into 3.Illustrations. Good ones go into the library for reuse in later sessions.
• Repetitive collection, such as logos — in an environment that can control a browser, you can hand off the collecting and file naming. But usage terms differ by image, so a person must verify the source and license.
• Cases and figures — Have the AI find them on the web, and require it to cite the source on the slide. I put “never use unverified figures” into the absolute rules.
• Always up to a person — deciding what to teach (content), judging which option is better (style), and the final review. None of this got smaller.

Before you try this · Four security rules

Connecting a folder to AI means showing it everything inside. Sort out these four things before you start.

1. Limit the connection to a single folder. Do not connect your entire home directory or cloud drive. If you connect only the working folder, any incident stays contained within it.
2. Keep client materials out of that folder. Never put files containing other people’s information in it: contracts, price lists, trainee rosters, and unreleased materials. You don’t need them to create training materials.
3. Build a confirmation step into anything that deletes. As in the command dictionary above, write “review first, move after approval” into the rules. It’s safer to default to moving files rather than deleting them.
4. A person gives outgoing materials a final review. Before distribution, have someone check the sources of any figures the AI pulled in, the usage terms of images, and whether any internal information slipped in.

What was left at the end of the day

That day’s output was two sets of course materials (a 16-slide special lecture and a 70-slide, 4-hour course). From the same content came a curriculum page for the website and the outline for this column, so one piece of work branched into three.

But the most valuable things were something else: the guidelines file and the illustration library. Course materials end with that course, but these two keep working. Tools change, but these assets stay. My takeaway: invest in assets, not tools.

Looking back, it was the same story I teach in the B2A course. A manual for people, execution rules for AI — one and the same document. In my last column I wrote, “Documents come before tools,” and applying that to my own work produced this.

FAQ

What exactly is a policy file?

It is a single text document that records my standards. I use the .md extension, but a Notepad file works too. You write rules one line at a time, such as “create new materials by duplicating this template” or “vertically center the text inside tables,” and have the AI read this document before it starts work. It works like the manual you would hand a new hire. The only difference is that the AI reads all of it, every time.

Do I need to know how to code?

No. Everything I did in this process was writing documents and putting files into folders. If you can write a rule as a sentence, that is enough. I am not a developer.

Does this work for tasks other than PPT?

Yes. The principle is the same for any work that has a set format plus recurring judgment criteria. Document work with a fixed form, such as proposals, reports, quotes, and meeting minutes, is an especially good fit, because you already have the template files and finished samples to start from.

I suggest doing just one thing today. Pick one recurring task, and every time you give AI a revision instruction, add it to a single document. In a week you’ll have about ten lines. That document is your organization’s first guidelines file. It’s much faster than shopping for a new tool.

We cover how to build this system and the B2A transition strategy hands-on in the B2A course (4 hours). If you are considering in-house training, feel free to reach out through the contact page.