Prompting Claude well

Every AI responds a bit differently, and Claude has specific things it responds to especially well — verified in Anthropic's own guidance. Learning these makes your results noticeably better. Good prompting is just clear communication, with a few Claude-flavored techniques.

Be clear, direct, and specific. Tell Claude exactly what you want, in what format, at what length. 'Write a 3-sentence, friendly reply declining this meeting' beats 'reply to this.' Specificity is the foundation for any AI, Claude included.

Use structure — including XML-style tags (Claude's signature tip). This is the most Claude-specific power technique. When your prompt mixes different parts — instructions, context, examples, and the document to work on — Anthropic recommends separating them with XML-style tags like <instructions>, <context>, <example>, and <document>. This isn't a hack; it's how Claude was designed to parse input, and it dramatically reduces confusion between your instructions and the material. For casual chat you don't need tags, but for anything complex — especially working over documents — wrapping the parts in clear tags noticeably improves results.

Give it a role and examples. Telling Claude who to be ('You are a careful copy editor') steers its tone and approach. And giving examples of the input-output you want (one to three) is one of the highest-leverage techniques — Anthropic strongly recommends it. Examples convey a pattern better than description.

Give it the documents. For anything based on source material, give Claude the material — paste or attach it — and tell it to base its answer on that. For multiple documents, wrap each in <document> tags. Grounding Claude in real source material makes it accurate and reduces made-up answers.

Let it say 'I don't know.' A verified Claude-specific tip: explicitly allowing Claude to say it's unsure measurably reduces hallucination. Add 'if you're not sure, say so' to prompts where accuracy matters.

The takeaway: Claude responds especially well to a few techniques (from Anthropic's own guidance). Be clear, direct, and specific (say exactly what, what format, what length). Use structure — Claude's signature technique is XML-style tags (<instructions>, <context>, <example>, <document>) to separate the parts of a complex prompt, which is how Claude was designed to parse input and dramatically reduces confusion. Give it a role and examples (one to three examples of the pattern you want is one of the highest-leverage moves). Give it the documents to ground answers in real material (wrap multiple in <document> tags). And explicitly let it say 'I don't know,' which measurably reduces hallucination. Casual chat needs none of this, but for complex work these make Claude noticeably better.