How this picture was made

Some pictures on this site were made with AI, and each one says so. They are here to explain an idea, not to decorate. This page shows how two of them were made, including the tries that went wrong, because the wrong ones teach more than the final picture does.

The two ideas

• What you let it touch: the AI in the middle, connected to your email, your calendar and your files. Every line is a door you opened.
• Why it forgets mid-chat: a long conversation slides through a window of fixed size. The newest messages are in view; the oldest fade out of reach.

First try: polished, and wrong

The first tool made two good-looking pictures. Both were wrong. The memory picture put the faded, forgotten messages inside the window, so it shows the AI holding on to what it actually loses. The other picture glowed neon green everywhere and drew the calendar as a notepad stuck inside an envelope.

Second try: a different tool, different mistakes

The second tool ignored the dark background, invented an envelope with calendar rings, and added a notepad. Asked for eight chat bubbles, it drew seven, with no window and no fade. Its one strength: it hands back a drawing you can edit, not just a flat image.

Third try: the keepers

The third tool got both ideas right on the first go. Still checked: the lines stop just short of the objects and the calendar has an odd number of days. Neither changes the meaning, so both pictures stayed.

What made the difference

1. Say exactly what is in the picture, and how many. "Exactly four objects" and "exactly eight bubbles" did more than any style word.
2. Give exact colours, as codes, and say which parts get the accent colour and which don't.
3. List what you don't want: no glow, no shadow, no text, no people. Image tools add all four unless told not to.
4. Keep words out of the picture. Image tools still garble text. Put labels on the page instead, where they can be read, searched and translated.
5. Try more than one tool. They differ far more than their demos suggest.
6. Check the result against the idea it is meant to teach, then decide. Two out of three tools failed that check while looking fine.

A flat, minimal line-art diagram on a solid near-black background ([background colour code]). One straight horizontal row of exactly [number] identical blank rounded speech bubbles, evenly spaced, left to right, thin 2px outlines. A bright [accent colour] ([accent colour code]) rectangular frame outline encloses only the [number] rightmost bubbles; those are crisp off-white. The bubbles to the left of the frame, outside it, fade step by step, the leftmost barely visible against the background. No glow, no shadow, no gradient fills, no texture, no 3D, no people, no text, no letters, no numbers anywhere. Generous empty space, calm editorial style.

The rules every AI picture here follows

• It explains an idea. If a real photo is the point, such as a session or a room, it's a real photo.
• It says "Made with AI" and names the tool.
• It never shows a real person, and never pretends to be a photo of something that happened.
• Labels are real text on the page, not baked into the image.

The same habits work for anything you ask AI to make: say exactly what you want, try more than one version, and check the result against what it was for. See Confident is not correct and Check against the goal.

Goes with the $25 webinar on Tuesday, November 10, 2026: When AI is wrong, and how to tell. Its reading list is in the Library.