Delimiters & Tags

Why <context>…</context> tags beat quotes. The tokenizer-level reason your structured prompts get ignored.

THE MINDSET SHIFT

“Claude was trained on XML-tagged documents. When you use <context> tags, you're not making the prompt prettier — you're speaking the model's native language.”
— SHE · YOUR AI GUIDE

Tokenizers split text into chunks the model attends to. Quotes ("like this") collapse into the surrounding text — the model treats the boundary as porous. XML-style tags create a hard attention boundary the model can reference ("as shown in the <context> block above…").

Anthropic explicitly documents this in their prompt engineering guide: use XML tags to separate context from instructions. OpenAI's structured-output mode does the same thing for JSON. The lesson generalizes — your separators should be unambiguous to the tokenizer.

“Claude 3 models follow XML-tagged instructions 18-22% more reliably than quote-delimited ones.”

Anthropic prompt engineering guide, 2024

“Models attend to tagged content as distinct attention blocks vs. unbroken continuous text.”

Tokenizer attention studies, 2023