Build Claude prompts with clean XML structure

Why XML tags work so well with Claude

A prompt often mixes very different things: instructions, background, a document to analyze, an example of the style you like. In plain text, Claude has to guess where each one starts and ends. Tags remove the guesswork. Anthropic's own documentation explains that wrapping each type of content in its own tag reduces misinterpretation, especially when a prompt combines instructions, context, examples and inputs. The tag names are not magic keywords; what matters is that they are descriptive and used consistently.

An Advanced-level prompt from this page looks like this:

<role>
You are a meticulous research assistant.
</role>

<context>
I am preparing a briefing for our board on remote work policies.
</context>

<task>
Summarize the attached survey results and highlight risks.
</task>

<requirements>
- Separate facts from opinions
- Summarize key points first, details after
- Flag anything that should be verified with a primary source
</requirements>

<output_format>
A bullet list, one idea per bullet.
</output_format>

Working with long documents

Claude is often used for reading long material: contracts, reports, transcripts, codebases. Anthropic's guidance for long inputs is specific: put the long documents near the top of the prompt, above your instructions and question, and wrap each one in its own tag with its source. For example, paste a report inside <document> tags before the generated prompt, then keep the task and requirements at the end.

A second technique from the same guide is to ask Claude to first pull out the relevant quotes, inside <quotes> tags, and only then answer based on them. This keeps the answer anchored to what the document actually says, which matters for summaries you will forward to others.

Explain the why, and say what to do

Claude responds well to reasons. "Keep each point under 20 words" works; "keep each point under 20 words because the summary will be read on a phone" works better, because Claude can apply the same reasoning to cases your rule did not cover. Add reasons to the <context> tag or next to a requirement.

Phrase rules as what to do rather than what to avoid. "Write in flowing paragraphs" is clearer than "do not use bullet points". In Expert mode, the <example> tag is where a short sample of your preferred style goes; Anthropic notes that examples are one of the most reliable ways to steer format and tone, and that several varied examples work better than one.

Projects, system prompts and reusing your structure

If you repeat the same kind of task, move the stable parts into project instructions or a system prompt and keep only the task in each message. The system prompt generator also outputs XML tags when you choose Claude as the target. Want to compare formats? The same idea becomes Markdown headings on the ChatGPT prompt generator and labeled sections on the Gemini prompt generator.