March 1, 2026
Return of the Angle Brackets
Why XML Tags Are So Fundamental to Claude
XML Is Back?! Claude’s ‘secret sauce’ sparks angle‑bracket brawl
TLDR: A viral post says Claude works best when prompts use XML-like tags to clearly separate instructions from content, claiming this was baked into training. Commenters are split between “structure helps” and “this is busywork,” with accusations of AI-written docs and nods that XML still runs under the hood — machine clarity vs. human simplicity.
XML, the 1998 internet relic with pointy brackets, just crashed the modern AI party — and the comments are losing it. A buzzy piece argues Claude, Anthropic’s chatbot, treats XML tags as VIP passes, using them in prompts and even training to mark what’s instruction vs. content, like quotation marks for computers. The claim: clearer “delimiters” mean fewer oops-moments, like that AWS prompt example where “Yo Claude” ended up in the email rewrite. Fans say structuring prompts with tags genuinely boosts results. Skeptics say… slow your roll.
That skepticism is spicy. One top comment calls the “Structure Prompts with XML” screenshot “screams AI‑written” and asks why anyone would trust “hallucinated documentation” about how to stop hallucinations. Another insists XML is still lurking under the hood anyway, noting Anthropic’s tool calls started as XML before they offered a JSON API — translation: the machine cares about brackets even if you don’t. A third camp wants product fixes: one user begs for four chat boxes — task, context, constraints, output — and even split results into “thinking” vs “content.” But the minimalist crowd fires back: wasn’t the whole point to give fewer instructions and let Claude figure it out? The joke of the day: “1998 called and wants its tags back,” as users argue whether this is smart structure or angle‑bracket cosplay. Meanwhile, Claude’s docs quietly wink, like, yes, delimiters matter — whether you write them or we do.
Key Points
- •The article argues Claude performs better when prompts are structured with XML tags.
- •It claims Anthropic incorporates XML tags in both inference-time prompts and training data for Claude.
- •Claude’s API documentation is described as promoting clear formatting rules and structured prompting with XML.
- •An AWS prompt engineering example is cited to show misinterpretation without explicit delimiters.
- •The article posits that paired delimiters enable transitions between layers of expression, aiding accurate model interpretation.