Wrap your prompts in <xml> tags so Claude can tell instructions from inputs

Claude was trained on prompts that use XML-style tags as structural delimiters. Wrapping pieces of a prompt in named tags — <document>, <instructions>, <example>, <format> — tells the model which span is reference material, which is the instruction to follow, and which is the worked example to imitate. It is the prompt-engineering equivalent of typing your variables: nothing forces you to do it, the program runs without it, but the second the prompt gets non-trivial you start saving yourself bugs.

Why this matters more in 2026 than it used to. Prompts now routinely carry multiple documents, prior turns, retrieved chunks, tool outputs, and a final task on top. When the structure is just a wall of prose, the model has to guess which sentence is your instruction and which is a quoted line from one of the attached files. With XML tags the boundary is explicit, and Anthropic's own prompt guide is direct about this: tagged sections improve accuracy on complex prompts, particularly when there are several inputs the model has to keep separate.

Separate the input from the instruction. The classic shape is <document>…raw text…</document> followed by <instructions>Summarize the document above in three bullet points.</instructions>. Same content as a flat paragraph, dramatically less ambiguity. Claude will not accidentally treat a sentence inside the document as a command to follow — a real failure mode when you paste in support tickets, emails, or model output that itself contains imperatives.

Give multiple inputs distinguishable names. If you are comparing two contracts, put them in <contract_a> and <contract_b>, not in two unlabelled blocks. Then your instruction can refer to them by name: “List clauses present in contract_b but missing from contract_a.” The model now has an unambiguous handle for each side, and you can chain follow-up turns without re-pasting either document.

Wrap examples in <example> so the format is unmistakable. For few-shot prompting, each example should sit inside its own <example> block, often with nested <input> and <output> tags. This is how Anthropic's own evaluation prompts are structured, and it shows: the model picks up the pattern from the shape of the tags, not from a heuristic about what looks like an example versus what looks like the real task.

The trap to avoid is treating XML tags as magic. They are not validated, they have no schema, and Claude does not parse them as XML — it treats them as strong delimiters. So do not bother closing tags with perfect whitespace, do not import an XML library, and do not invent a 14-tag taxonomy when three would do. Pick a small vocabulary, use it consistently across a project, and put it in your CLAUDE.md so every prompt in that repo speaks the same dialect.

One more reason to adopt this now: tagged prompts are easier to template, easier to diff in version control, and easier to read six months later when you have forgotten which paragraph was the instruction. See the try-it block for a side-by-side you can run in the Claude app or any SDK call.