A developer wraps every sentence in a prompt with its own XML tag: <instruction1>, <instruction2>, <context1>, <context2>... The prompt has 20 separate tags. What is the problem?
Answer: Over-tagging fragments the prompt and degrades model performance — tags should mark major semantic components, not individual sentences
XML tags are effective for marking major semantic components (instructions, context, examples, query). When every sentence gets its own tag, the tags stop communicating semantic meaning — the model can't distinguish 'this is a separate major component' from 'this is an individual sentence.' Over-tagging also fragments the logical flow the model uses to understand relationships between sentences. There is no API limit on tag count (A). Numbered XML tags are syntactically valid (C). Claude doesn't repeat tagged content in output (D).