XML-теги и разделители в подсказках

Prompts often mix several kinds of text: your instructions, material to analyze, examples of the output you want, and variable input from users. Without clear boundaries, a model can blur them. It might treat a sentence inside a pasted email as an instruction, continue an example instead of answering, or summarize your instructions along with the document. Delimiters fix this by labeling each part. Anthropic's documentation recommends XML tags in prompts for Claude. OpenAI's prompting guidance similarly suggests delimiters such as triple quotes, Markdown headings or XML to mark separate sections. There is no official list of tag names. <contract> or <customer_email> works as well as <document>, as long as you use the names consistently and refer to them in your instructions, for example "Using the report in <report> tags..." Models have seen large amounts of HTML and XML during training, so tagged structure is familiar to them. Tags help in both directions: - On input, they separate content. - On output, asking for <answer> or <json> tags gives your code a reliable place to find results. Putting reasoning in one tag and the final answer in another keeps the two apart. Nesting expresses hierarchy. For example, a <documents> tag can hold several <document> elements, each with its own <source> and <content>. A common misconception is that tags are a security boundary. They make your intent clearer and can reduce accidental instruction-following. But text inside a tag can still contain a prompt-injection attempt, and a determined attack may still succeed. Treat delimiters as clarity tools and combine them with other defenses, such as limited tool permissions and checks on the output. Another misconception is that the XML must be strict and valid. Models handle informal tags well. Still, matching opening and closing tags consistently avoids ambiguity.