Anthropic XML Tags Cheat Sheet

A compact reference for using XML-style prompt tags with Claude.

Canonical skeleton

<instructions>
State exactly what Claude should do.
</instructions>

<context>
Background facts, constraints, or domain assumptions.
</context>

<examples>
  <example>
    <input>Example input</input>
    <output>Example output</output>
  </example>
</examples>

<input>
The live task or source material.
</input>

Recommended tag roles

<instructions> — operative directions
<context> — non-operative background
<examples> — container for demonstrations
<example> — one demonstration pair or case
<input> — the live item Claude should process
<documents> / <document> — repeated source documents

The exact names are flexible. Anthropic's stable recommendation is to use descriptive tags consistently.

Good default patterns

Single task

<instructions>
Summarize the argument in 5 bullets.
</instructions>

<input>
...
</input>

Task with background constraints

<instructions>
Extract safety-relevant claims only.
</instructions>

<context>
Ignore marketing language and rhetorical flourishes.
</context>

<input>
...
</input>

Few-shot prompting

<instructions>
Rewrite each passage in plain language.
</instructions>

<examples>
  <example>
    <input>The physician commenced the procedure.</input>
    <output>The doctor began the procedure.</output>
  </example>
</examples>

<input>
The committee rendered its determination.
</input>

Multi-document prompting

<instructions>
Compare the documents and list disagreements.
</instructions>

<documents>
  <document index="1">
    ...
  </document>
  <document index="2">
    ...
  </document>
</documents>

Do / don't

keep each region semantically distinct
use nested tags when the source material is naturally nested
separate examples from live input
keep tag names stable throughout the prompt

Don't

rely on tags to replace unclear instructions
mix demonstrations and live data in the same block
treat this as a strict parser contract
assume the prompt must be valid XML

Mental model

Think of XML tags here as legibility markup for the model, not as a machine-validated schema. The goal is to make prompt boundaries obvious so Claude does not confuse instructions, context, examples, and live input.