Why wrapping the same text in XML changes the answer Claude gives

Paste a paragraph into Claude as plain text and ask for a summary, and the result comes back generic. Paste the exact same words again, this time wrapped in XML tags, and the summary suddenly gets sharp, pulling out the details that matter. Nothing about the content changed. Only its shape did. That's strange enough to be worth explaining.

The puzzle: same words, different answer

If the words are identical, logic suggests the answer should be too. But formatting alone shifts the quality of what Claude returns, and that shift has nothing to do with adding or removing information. The content stays fixed; only the structure wrapped around it changes.

The anchor: one product description

Take a single product description as a test case. Paste it in as plain text, and Claude's summary comes back generic, the kind of output that could apply to a lot of different products. Wrap the exact same words in XML tags, marking off sections like title, content, and category, and the summary gets sharp. It pulls the right details out specifically because the structure told Claude where one piece of information ends and another begins.

Why: trained on the shape

The explanation isn't about a trick or a hidden feature. Claude was trained on enormous amounts of consistently tagged, structured text, documents where fields like title, content, and category consistently sat inside predictable markup. That training built a strong expectation for those shapes. Text that matches the structure Claude learned to expect gets processed more coherently. Text that doesn't match that shape gets processed less well, even when the underlying content is exactly the same. Wrapping text in XML isn't a trick for getting better answers. It's closer to speaking the shape Claude was already trained to expect.

The anchor returns: Robot Building Kit

The difference is easiest to see side by side. As plain text, a product entry might read simply as "Robot Building Kit, 200 pieces," with no clear boundary between the name and the description. Tagged, the same information gets split into distinct fields, one for title, one for content, one for category, each wrapped in its own markup. The words are identical in both versions, but the tagged version tells Claude exactly where the title ends and the description begins, because that boundary matches a shape it saw constantly during training.

What this does and doesn't mean

It's worth being precise about the claim here. This isn't a statement that XML is the only structure that works, and it isn't a claim about tokenization or attention mechanisms inside the model. The explanation stays at the level of training-distribution expectation: Claude processes input more coherently when it matches patterns common in its training data, and consistently tagged XML is one very common pattern. It's also not a claim that every prompt should always be tagged. It's a specific, testable effect worth checking on your own text.

Key takeaways

• The same words can produce a noticeably different quality of answer from Claude depending on whether they're plain text or wrapped in XML.
• Claude was trained on huge volumes of consistently tagged, structured text, which built a strong expectation for that shape.
• XML tags work by marking clear boundaries, like where a title ends and a description begins, that match patterns Claude saw constantly in training.
• This is a training-distribution effect, not a claim about tokenization, attention internals, or XML being the only structure that helps.

Try it yourself

Open a Claude conversation and paste in any short paragraph of product or document text, once as plain text, and ask for a summary. Then paste the identical text again, wrapped in XML tags, title, content, category, or whatever fields fit, and ask for the same summary. Compare the two answers on your own text rather than taking the claim on faith. This walkthrough is part of the Claude Basics playlist that Humanitarians AI produces to explain practical prompt-structure ideas like this one.