Write alt text for what the image is for, not what it looks like

Upload a photo to an AI tool and ask for alt text and you'll get a description: "A group of four people sitting around a table with laptops in a bright office." Accurate, and often wrong for the page, because alt text isn't a description of the picture. It's what the picture is there to tell someone who can't see it (WCAG 1.1.1).

A team photo on the "About us" page might need "The clinic team, outside the Sheffield site". The same photo on a recruitment page might need nothing at all if the text beside it says everything. A chart needs the finding, and the data in text nearby.

So tell the model what the image is for.

Write alt text for the attached image.
Where it appears: [PAGE AND SECTION]
What it's there to do: [e.g. show the entrance so people can find it /
back up the claim in the paragraph / decoration only]
Text next to it: [PASTE]
Rules:
- One sentence, under 125 characters if possible.
- Don't start with "image of" or "picture of".
- Include any text that's in the image.
- Don't repeat what the caption or the paragraph next to it already says.
- If the image is only decoration, say "DECORATIVE: use empty alt".
- If it's a chart or diagram, say so, and tell me a long description is
  needed instead, and what it should cover.
Give me two options.

Then check it. Models misread faces, signs, small text and what a chart is actually showing, and they'll describe a person's age, mood or background with a confidence they haven't earned. Say only what's relevant to the page.

A chart's real alt text is a short statement of what it shows, with the numbers in a table or the surrounding text. Don't leave that to the model.

The test is simple. Read the page out loud with the alt text and without the image. Does it still make sense?