It means wrapping each part of your message in a simple named label, so Claude knows what each part is for.

A tag is a short word placed between angle brackets, like a label at the start of a section, with a matching label at the end to close it. Everything between the opening label and the closing label belongs to that section. For example, you might place a pasted email between a “email” label and its closing label, and your request between an “instructions” label and its closing label. Claude can then see at a glance which part is the email to be read and which part is the task to be done.

This style of labeling is called XML-style tagging, because it borrows its look from a long-standing format used to organize data. You do not need to know anything about that format to use the idea. All you need is to open a label, write your content, and close the label. There are no strict rules about which words you may use, and the labels do not have to come from an official list.

flowchart LR
    A["Messy block of text"] --> B["Add labels around each part"]
    B --> C["Instructions"]
    B --> D["Document"]
    B --> E["Examples"]
    B --> F["Output format"]
    C --> G["Claude sees clear sections"]
    D --> G
    E --> G
    F --> G
        

A few simple words appear again and again in this topic. Here they are in plain English.

Plain, unlabeled prompts work for small jobs but start to fail as they grow.

For a short, simple request such as “Explain how rainbows form”, no organization is needed. But real work is rarely that tidy. You might want Claude to read a long report, follow a set of rules, match the style of a sample, and then answer in a particular layout. Putting all of that into one stream of sentences creates several risks.

Tags give Claude clear signals about what each part of your message is.

Anthropic’s guidance says that when a prompt has several components, such as context, instructions, and examples, tags help Claude read it more accurately and tend to improve the quality of the answer. A tip in the official documentation suggests using labels such as instructions, example, and formatting to separate the different parts, which prevents Claude from mixing up instructions with examples or context.

The official page highlights a few benefits. It points to clarity, because each part is clearly separated. It points to accuracy, because there is less chance of parts being misread. It points to flexibility, because you can find, add, remove, or change one part without rewriting everything. And it points to parseability, because if Claude uses tags in its reply, a program or a person can pull out exactly the piece needed.

The basic pattern is the same every time: open, fill, close.

You can invent your own, but these common labels cover most everyday needs.

Let us start with the most common use: separating your request from the material you want Claude to read.

Suppose a customer sent a long, rambling email and you want a short summary plus a suggested reply. Without labels, you might paste the email and write your request right under it, and the two could run together. With labels, the structure is crisp.

When you share more than one document, tags keep them from blurring into one.

Imagine you are comparing two job offers. You paste the text of both and ask Claude which has the better terms. If the texts are simply pasted one after another, Claude may struggle to tell where the first offer ends and the second begins, particularly if they use similar wording.

flowchart TD
    A["Whole prompt"] --> B["documents"]
    A --> C["instructions"]
    B --> D["offer_a"]
    B --> E["offer_b"]
    D --> F["source"]
    D --> G["content"]
    E --> H["source"]
    E --> I["content"]
        

Tags are especially useful when your prompt includes sample answers.

Samples are a powerful way to show Claude the style you want, but they must be clearly marked so they are not mistaken for the real task. Tags solve this neatly. Place the samples inside an “examples” section, with each sample wrapped in its own “example” label, and keep your real input in a separate “input” section.

If you would like to learn more about how samples help, the idea of showing an example in your prompt is a technique on its own. Tags and examples work wonderfully together, because tags keep the samples tidy and clearly separate from the real job.

Tags can organize Claude’s answer too, not only your input.

You are not limited to labeling what you send. You can also ask Claude to label parts of what it sends back. This is helpful when your answer has several pieces, such as a short summary and a longer explanation, and you want to find one quickly or pass it into another tool.

sequenceDiagram
    participant You
    participant Claude
    You->>Claude: Task, material in labeled sections, and the output labels you want
    Claude-->>You: Reply with reasoning and answer in separate labeled sections
    You->>You: Copy only the answer section you need
        

When content has natural layers, place tags inside other tags.

Some information is layered. A list of customer reviews contains many separate reviews. A collection of documents contains several documents, each with a source and a body. Nesting means placing smaller labeled sections inside a larger labeled section, which reflects that structure.

Anthropic’s guidance mentions nesting as a good practice. The key rule is to close the inner tags before closing the outer one, just like boxes that must be closed in the right order. An outer “reviews” label might hold several inner “review” labels, and the outer closing label comes only after all the inner ones are closed.

Where you put each labeled section can also affect the quality of the result.

Guidance for working with long material suggests putting long documents near the top of your prompt, and placing your question or instructions after them, near the end. Many people find this order works well, especially with large amounts of text, because Claude reads the material first and then sees exactly what you want done with it.

flowchart TD
    A["1. Documents"] --> B["2. Context and examples"]
    B --> C["3. Instructions and question"]
    C --> D["Claude answers"]
        

Tags are not the only tool. Here is how they compare with other common methods.

These methods can be used together. You might use numbered steps inside your instructions section, while using tags to separate the instructions from the document. Think of tags as the outer structure and the other methods as tidying inside each section.

Tags are helpful, but a short simple message does not need them.

flowchart TD
    A["How many parts does your message have?"] --> B{"More than two or three kinds of content?"}
    B -- "Yes" --> C["Use tags"]
    B -- "No" --> D{"Could one part be confused with another?"}
    D -- "Yes" --> C
    D -- "No" --> E["Plain writing is fine"]
        

The same idea works wherever you write to Claude.

In the regular Claude chat, you simply type or paste labeled sections into your message. There is nothing to switch on. Tags are only text, and Claude reads them like any other part of your message.

In ongoing work such as Projects, where you save instructions and background for repeated use, tags help you keep those saved instructions neat: one section for the role and tone, another for the rules, another for background facts. Because they are labeled, you can update one section later without rewriting the others.

Developers who build their own tools with the Claude API often write a system prompt, which is a standing set of instructions given to Claude before the conversation begins. Tags are very popular there, because the same structure can be filled with new content each time while the labels stay constant.

Labels organize content. They do not make it safe.

It is tempting to think that wrapping a pasted web page or email in a “document” label makes anything inside it harmless. That is not quite true. Tags tell Claude what the content is, but a pasted text can still contain sentences written to push an AI into doing something. These are sometimes called prompt injection attempts, where hidden or sneaky instructions are placed inside material the AI reads.

Labeling the material as a document helps Claude treat it as something to read, not as orders from you. But you should still be careful with text from unknown sources, especially if Claude can take actions for you, such as sending messages or using connected apps. Review what Claude proposes to do before anything important or irreversible happens.

A handful of habits cause most problems with tagged prompts.

Tags bring real benefits at a small cost in effort.

Keep these habits in mind whenever you organize input with tags.

Quick answers to the questions beginners ask most often.

They are short named labels placed around parts of your message, with an opening and a closing label, so Claude can tell what each part is, such as instructions, a document, or an example.

No. You only need to open a label, write your content, and close the label. The style looks like XML, but you do not need any technical knowledge.

No. Any sensible, descriptive name works. What matters is that the name fits the content and that you use it consistently.

Whenever your message mixes several kinds of content, such as instructions, a pasted document, examples, and a requested format. For short, simple requests they are not needed.

Yes. Name the labels in your instruction, such as asking for reasoning in one labeled section and the final answer in another. This makes it easy to pick out the part you need.

It is a good habit. A closing label clearly marks the end of a section and keeps the structure clean, especially when tags are nested.

Yes. This is called nesting and is useful when content has layers, such as a list of documents that each have a source and a body. Close the inner tags before the outer one.

Many people place long documents first and the instructions or question last, so the task is clear and fresh at the end. Try the order that makes your prompt clearest.

Yes. They are simply text in your message, so there is nothing to turn on. They also work in saved project instructions and in system prompts for developers.

No. Tags organize content, but text from unknown sources can still contain misleading instructions. Treat such material as information only and review any action before approving it.

They solve a different problem. Tags give each section a name and a clear end, which is especially useful for separating documents, examples, and instructions. You can combine them with headings and lists inside each section.

Anthropic’s prompt engineering documentation has a page on using XML tags to structure prompts, which explains the benefits and best practices in more depth.

Tags are simple named labels that you place around parts of your message to Claude, so it can tell instructions from documents, examples, and formatting requests. They work like labels on moving boxes: they make everything easy to find and prevent parts from being mixed up. Anthropic’s own guidance recommends them for prompts that combine context, instructions, and examples, and highlights clarity, accuracy, flexibility, and easy extraction as the main benefits. You choose descriptive names, open and close each tag, refer to the labels in your instructions, nest them only when content truly has layers, and put long material first with the task at the end. You can also ask Claude to label its reply. Tags are unnecessary for short, simple messages, and they do not make untrusted content safe, so keep reviewing important actions and checking important facts.