The Practice Room / Your AI ___ Sucks / Episode 2
Why your AI invitation looks cheap
The short answer
It is almost never the information. The date is right, the address is right, and it still looks like everybody else's.
It looks cheap because you asked for the finished thing in one prompt. A designer never does that. A designer builds it in parts and stacks them: the picture, the container, the words.
Why one prompt cannot work
When you ask for a whole invitation at once, everything arrives fused. The words are painted into the picture, so they come back bent and misspelled and you cannot move them. There is no separate layer to adjust, which means the only way to change anything is to generate the whole thing again and hope.
ONE ASK AGAINST THREE
ASKED FOR THE WHOLE THING
- Letters that are nearly words
- Type you cannot move or correct
- A picture welded to its text
- A change means starting over
ASKED FOR THE PARTS
- Words set by you, spelled right
- Every element still movable
- A picture that stays a picture
- A change is one value
The three layers
PICTURE, CONTAINER, WORDS
Layer two is the one people skip, and it is the one that stops the result looking like text dumped on a stock photo.
Do it yourself
- Open Command Prompt, move into your folder with cd, type your tool's name and press enter.
- Ask for the picture only. Describe the scene and say plainly that there should be no words anywhere in it. Ask for the middle to be left calm, because that is where the type is going.
- Place the container. A shape with an edge, sitting over the picture, still empty. This is what gives the words somewhere to live.
- Set the words last. Type them yourself, or place them with code, so they are straight and spelled correctly.
- Change one thing and look again. Because it is stacked, you can.
Hierarchy, so it does not read as a caption
Four lines at the same size is not a design, it is a paragraph. Give each tier a different job, and change more than the size: change the colour and the case too.
FOUR TIERS, FOUR TREATMENTS
WORKS
- WHAT it is, biggest by far
- WHEN, second, and the only accent colour
- WHERE, clearly smaller and plain
- The ask, smallest and quietest
DOES NOT
- Four lines within a few points of each other
- Everything in one colour
- A gradient running through small type
- Pale text on a pale background
Check yourself
Why did my words come out misspelled?
They were generated inside the picture. Image models paint letter shapes, they do not type. Ask for a picture with no words at all, then add the words as their own layer.
What counts as a container?
Any shape that holds the text: a panel, a card, an arch, a scalloped frame, even a simple keyline with a darkened area behind it. Its job is to give the words contrast and a home.
Should the type be a gradient?
No. A gradient through small type lowers contrast exactly where it matters most and is harder to read for anyone with low vision. Keep gradients for the decorative marks and set the words in one solid tone.
When it goes wrong
- The text disappears into the picture. You skipped layer two. Put a container behind it.
- It looks like a template. Usually the picture is generic. Describe the room, the light and the objects, not the occasion.
- You cannot fix one line. The words were generated rather than set. Rebuild with the words as their own layer.
What to carry forward
The artefact changes and the method does not. A flyer, a menu, a poster, an infographic, a book cover: picture, container, words. Ask for the parts, never the finished thing.
Words used here
- Can ChatGPT generate images
- How do I generate images with AI for free
- Who owns an AI generated image
Go deeper
What she says in the film
Most people ask AI for the whole invitation at once. That is the only reason it comes out looking cheap. It is never the information. The date is right, the address is right, and it still looks like everybody else’s. A designer does not work that way. A designer builds it in parts and stacks them. So I move into my folder with cd, I type codex, and I press enter. Layer one is the picture, and I ask for it with no words in it at all. A winter table, candles, snow through the window, and the middle left calm. No words in that first ask. Words generated inside a picture come out bent and misspelled, every time. Layer two is the container. That green shape with the gold edge, placed on its own, still empty. It gives the words somewhere to sit, and it is what stops the whole thing looking like a stock photo with text dumped on it. Layer three is the words, and they go on last, where I control them. So they are straight, and they are spelled right. Picture. Container. Words. And once it is stacked like that, changing it is one word. Gold tonight. Silver tomorrow. Same invitation, nobody starting over. It works the same for a flyer, a menu or an infographic. Ask for the parts. Subscribe. No permission required.
That is the last episode of Your AI ___ Sucks so far. New ones land as they are filmed.
Book a session and we do it together on a call.
Book an hourAsk what did not make sense
Comments go through our own form and are read before they appear. Links are not allowed; describe it in words.
3 more from the room.

Why do all AI flyers look the same?
Different party, different month, same flyer. It got every detail right, and that is exactly why a better prom…

How to make a flyer with ChatGPT that looks designed
The layered build, start to finish, in the ChatGPT desktop app. A cream page, one photograph asked for once, a…

Is AI slop the machine’s fault?
Get the next walkthrough.
One email when something goes up. The argument, the film and the term, nothing else. One click to leave.
Have a question now? Ask Mstimaj, she knows the archive.