Try describing a slightly misaligned button in words. You'll end up writing three sentences about padding, flex alignment, and a color that's "a bit too dark," and the model will still guess wrong half the time. Now paste a screenshot next to the design mock instead. The gap is obvious in one glance — to you, and increasingly, to the model too.
The shift is already showing up in real releases
Anthropic's own description of Claude Fable 5 lists this directly as one of its core abilities: the model "uses vision to check outputs against goals," evaluating its own coding work against the original design or screenshot rather than just re-reading the code it wrote. That is a meaningfully different workflow than the classic loop of prompt, generate, read the diff, correct. It's prompt, generate, look at the actual rendered result, and self-correct against what it sees.
Why a picture can beat a paragraph here
A wall of text describing a UI state is lossy in a specific way: it forces you to notice and name every detail that matters, and you will miss some. A screenshot doesn't have that problem. It carries spacing, color, hierarchy, and layout all at once, without you having to consciously describe any of it. For visual and layout-heavy work, that's simply more signal per prompt, even before you get into token counts.
It's worth being precise about the token economics, because this gets oversimplified a lot. An image is not automatically cheaper than text in raw token terms — most vision models tokenize images into a meaningful chunk of context, sometimes more than the equivalent text description would cost. The real saving isn't always in tokens per message. It's in round trips. One screenshot plus "fix this" often replaces three or four back-and-forth messages of increasingly specific text corrections, and it's the avoided rounds that save time and money, not necessarily the single message.
Where this actually helps
- UI and layout bugs. Show the broken state next to the intended design instead of describing both.
- Self-verification loops. Let the agent take its own screenshot after a change and compare it to the goal before declaring the task done, the same pattern Fable 5 uses internally.
- Charts, diagrams, and messy scanned documents. Vision models parse table structure and nested layouts in files far more reliably than a text dump of the same content.
Where it doesn't help
Business logic, edge cases, and anything that lives in data rather than pixels are still better explained in words. A screenshot of a spreadsheet doesn't tell the model your reconciliation rules. Don't reach for an image just because you can — reach for one when the thing you're trying to communicate is fundamentally visual, and stick to clear text everywhere else.
