Claude adds XML-tagged prompts for task, context, constraints, and output format

• A Twitter thread summarizing Anthropic guidance says Claude responds better when briefs are split into XML-style sections rather than written as one conversational paragraph, with tags for task, context, constraints, and output format thread summary.
• The same walkthrough argues the gain comes from structure: in the tag explainer, XML tags act as separate containers, and the hierarchy example says nested tags help Claude weigh main goals against supporting details.
• For creative work, the most useful patterns are cleaner examples and stricter deliverables: the isolation demo separates good and bad references, while the validation example shows rules for length, bullets, and citations.
• Claims like 40% better output and 60% fewer hallucinations come from the thread's examples, not from independently verified benchmarks, though the model note says the approach applies across Haiku, Sonnet, and Opus.

The practical shift is simple: stop writing one long brief and start labeling each part of the job. The thread's starter template in the basic template uses four buckets — <task>, <context>, <constraints>, and <output_format> — which maps neatly to creative work like moodboards, shot lists, lyric rewrites, or scene treatments.

That structure is meant to tell Claude what each block is for, not just what words to parse. In the tag explainer, the author describes tags as separate context containers, and the hierarchy example adds that outer tags carry the main objective while nested tags hold audience, tone, or style details.

The clearest creator use case is reference handling. Instead of mixing inspiration and instructions in one paragraph, the isolation demo shows separate <good_example>, <bad_example>, and <your_task> blocks so the model can borrow the right pattern without copying the wrong one. That is useful when you want a trailer script in the rhythm of one sample, but not its clichés.

Constraint-heavy briefs are another good fit. According to the validation example, a dedicated <validation_rules> block can specify exact length, section count, and citation requirements; the before-after prompt pairs that idea with a tagged product-description prompt that the thread says outperformed a plain-English version.

This is a workflow recommendation, not a confirmed product launch, and the strongest claims in the thread are anecdotal. The post says XML prompting works across Claude models in the model note, but the quoted gains — better quality, stronger constraint-following, less hallucination — are examples from one creator's summary rather than published benchmark results. For creative teams, that makes the technique promising precisely because it is easy to test against your existing prompts.