Structure with XML tags

Real prompts mix things up: your instructions, some background, an example or two, and the text Claude should work on. Run it all together and Claude has to guess where one part ends and the next begins.

XML tags fix that. Wrap each kind of content in its own named tag, like <instructions>, <context> or <email>, and Claude can tell the parts apart. Anthropic recommends tags for exactly this: prompts that mix instructions, context, examples and inputs.

Here's why it matters. You ask for a one-line summary of an email, and the email happens to contain an instruction of its own.

On the left, the email's P.S. reads like one more request from you, so it gets answered. Current Claude models often spot this kind of thing anyway, but why make them guess? On the right, the tags make it clear which words are yours and which are just the text to summarise.

An honest caveat: tags make that line clearer, but they're not a security wall. Text inside a tag can still try to steer Claude, and sometimes it works. If you're handling untrusted text in a real app, tags are one layer of defence, not the whole plan.

A few habits make tags work well.

Name them for what's inside. <email>, <contract> and <notes> beat <text1>. Use the same names every time, and refer to them in your instructions: “Summarise the email in the email tags.”

Nest when there's a hierarchy. Several documents can each sit in a <document> tag inside one <documents> tag.

Ask for tagged output too. “Put the summary in <summary> tags” gives your code an easy thing to pull out, and keeps any chatter outside it.

Put long text first. With long documents, Anthropic finds that placing them at the top and your question at the end gives better answers.

Your turn. The draft runs everything together. Wrap the email in <email> tags and ask for the answer in <summary> tags. Then run it and see whether the P.S. gets obeyed.