Auditing Alt Text in InDesign, in One Pass
You finish the document. Then someone asks for an accessible PDF, and you remember: every image needs alt text. So you open Object → Object Export Options, pick an image, type a description, and do it again. And again. On a real document—a report, a catalogue, a magazine—that’s dozens of trips through the same dialog, which is exactly why it quietly gets skipped. Then an accessibility check flags every figure, and it’s no longer optional: under standards like WCAG and PDF/UA—and, increasingly, the law—those descriptions have to be there.
Lately it’s worse. Recent InDesign versions started auto-filling alt text with AI-generated descriptions, stamped with the words “AI generated.” They land on whatever you place—including the client’s photoshoot—and unless you catch them, they ship. Turning the feature off is its own chore; removing the placeholders is per-image drudgery all over again.
So you need two things at once: catch the images that are missing alt text and the ones carrying that “AI generated” placeholder, and replace them with proper descriptions. And the descriptions have to be right—which, for images, is trickier than it sounds.
A document that’s a realistic mess
We built a test document for exactly this and put it in the Sidekick examples repository. It’s a two-page, monochrome editorial piece with twelve photographs, in the state a real document tends to be in:
- five images carry Adobe’s “AI generated” placeholder,
- four have no alt text at all,
- three were already described properly by hand.
Nine problems to fix, three to leave alone.
But the interesting part is how five of the photos are placed. One is cropped tightly by its frame. One sits under a headline. One is partly covered by a solid block. One is rotated. One is clipped to an oval. These are the cases where automatic alt text goes wrong—because the right description isn’t of the photo file, it’s of what the reader actually sees on the page.
Running the audit from Claude Desktop
Open before.idml in InDesign, make sure Sidekick is connected in Claude Desktop, and ask:
Audit the alt text on every image in this InDesign document. Show me which images are missing alt text and which have an auto-generated “AI generated” placeholder, then write proper alt text for those—describing only what is actually visible in each image as it sits on the page—and leave the images that already have good alt text alone.
Claude reads each image’s current alt text, looks at the pages as they’re laid out, and reports back: the five AI-generated placeholders and four blanks, flagged; the three good ones, recognised and left alone. Then it writes real descriptions for the nine—none of them starting with “AI generated”—and writes them straight back into the document.
That’s the whole loop, on the whole document, from one sentence. Here it is running end to end:
Why “what’s visible” is the hard part
An image on a page is rarely the whole image file—it might be cropped by its frame, tucked under a headline, half-covered by a block, rotated, or masked to an oval. Because Claude can now see each image as it sits on the page, it describes the visible crop, not the raw asset on disk. And for alt text, that distinction is the whole game: a screen-reader user only gets what the page actually communicates, so describing the hidden half of a photo—or the empty corners around a rotated frame—doesn’t help them, it misleads them.
This is precisely where Adobe’s per-image AI—working from the raw file, blind to the layout—gets it wrong. Claude gets it right because it sees the page the way the reader does.
When it’s a judgment call
Not every image has one obvious description—an abstract texture, a photo where the subject is genuinely ambiguous, a picture that’s decorative in one context and load-bearing in another. For those, you can keep a human in the loop without slowing down the clear-cut ones. Add a line to the ask—“and show me the image before writing if a description is a judgment call”—and Claude presents it inline, you decide, and only then does it write. The images that are unambiguous still get handled in the same pass; you only get pulled in where your judgement actually changes the answer, which is exactly where accessibility needs a person.
From a good prompt to a reusable skill
The first time, you type that careful sentence. But the value isn’t the prompt—it’s the recipe underneath it: read the current alt text, flag the AI-generated and the missing, describe only what’s visible, leave the good ones, write it back to the right frames. That’s a fair amount to remember, and to get right, every time.
Which is the whole case for a skill. Package that recipe once, and “audit my alt text” becomes a single ask that does the right thing on every document—this one, the next one, and the one your colleague is fighting with. The nuances that took a paragraph to spell out are baked in; nobody has to re-derive them under deadline. It’s the difference between a clever one-off and a capability your whole team can lean on. (We walk through turning a session into a skill in From Session to Skill.)
Try it yourself
The example—the two-page document with its deliberate mix of missing, AI-generated, and good alt text—is in the Sidekick examples repository. Clone it, open before.idml, and hand Claude the prompt above.
The photos are freely-licensed and converted to greyscale; the type is Minion Pro, which ships with InDesign, so the file opens cleanly for anyone.
Requirements
- Adobe InDesign with the Sidekick plugin installed
- Claude Desktop (or any other MCP-compatible AI assistant) with Sidekick connected
- A document whose images need alt text—or whose alt text needs cleaning up
One image is a minute by hand. A whole document, before a compliance deadline, is the kind of afternoon you don’t get back—and the kind a single ask can hand back to you.