Ask an AI chatbot for a slide and you'll get something good in under a minute: a clean layout, a sensible hierarchy, the right kind of structure. Then you ask for the next slide, and the next deck, and you notice it never quite lands on brand. AI chat reliably gets a slide about 80% of the way. The last 20% is where a brand lives, and better prompts don't close that gap.
TL;DR
- AI chat is great at getting past the blank page. It's weak at consistency.
- The gap isn't the model's taste. A brand's rules live in someone's head, not in the prompt.
- The fix is structure: rules written down, layouts built from shared parts, and AI choosing rather than drawing.
What does the 80% look like?
I keep seeing the same pattern around me. Someone needs a slide, opens a chat window, describes the content and asks for a design.
What comes back is genuinely useful:
- Speed. A first layout in seconds instead of half an hour.
- Structure. Messy notes become a heading, three points and a takeaway.
- A starting point. For someone who isn't a designer, that alone changes the afternoon.
For one-off internal slides, 80% is often enough.
Where does the last 20% go?
It goes in the details only a designer notices, and every client notices without knowing why.
- No consistency. Ask twice and you get two layouts, two spacing systems and two type sizes. Nothing carries over to the next deck.
- Words that don't fit. Long copy gets squeezed until it's tiny, or it runs past the edge.
- Close, not right. The colours are correct but the rules aren't: headings that should be uppercase aren't, and the accent colour takes over the slide.
None of this is the AI being bad at design. It's being asked to guess rules nobody wrote down.
Can it improve?
Partly, and most of it works with tools you already have.
- Write brand rules as rules, not adjectives. "Clean and bold" can't be checked. "Headings uppercase, five columns maximum" can.
- Give the AI a menu, not a canvas. A template with clearly named layouts gives any AI something consistent to choose from.
- Never shrink to fit. When copy doesn't fit, split the slide or cut the words.
What prompting can't fix is consistency across people and across decks. A prompt holds an intention. It doesn't hold a standard.
What am I testing?
The answer I keep coming back to comes from web design: atomic design. Instead of designing whole pages, you design small parts and build everything from them. Atoms are the basics, like a text style, a rule line or a colour. Molecules combine them, like a numbered item or a stat with its label. Layouts are assembled from molecules. Change one part, and every layout that uses it updates.
I'm applying that to slides in a project I call APE, the Atomic Presentation Engine. It works in three layers:
- Tokens and rules in a file. Fonts, colours and spacing, plus the judgments: column limits, word budgets, how much accent a slide can carry.
- Layouts built from shared parts. Grids, timelines, tables and stats that adapt to their content and split onto a new slide when they run out of room.
- AI that chooses, code that draws. The AI reads the content and picks a layout from the menu. Code places every element, then checks the finished slide against the rules.
It's early. The layouts work, and the checks catch real errors, including a few in my own code. Next is the part where the AI reads a brief and plans the whole deck. Once it's tested against the alternatives, I'll write up what held and what didn't.
In the meantime, a question worth asking your own team: if your brand guidelines had to be tested, how many of them could actually be checked?
