The Best Claude Prompts (and How Claude Differs from ChatGPT)

Anthropic’s Claude models share the same prompting fundamentals as every other model — role, task, context, constraints, example — but a few Claude-specific habits consistently improve results. This guide covers those habits and gives you tested prompts to start from.

Use XML-style tags to structure input

Claude was trained to pay close attention to XML-style tags. When a prompt mixes instructions with a long document, wrapping each part in tags removes ambiguity:

<instructions>
Extract every action item and its owner from the meeting
notes. Return a markdown table with columns: Owner, Action,
Due date. If a due date is missing, write "TBD".
</instructions>

<meeting_notes>
{notes}
</meeting_notes>

This is the single most useful Claude-specific technique, and it scales: the more distinct pieces of input a prompt has, the more tags help.

Let Claude reason before answering

For analysis, maths, or multi-step decisions, ask Claude to work through its reasoning in a <thinking> block first, then give the final answer separately. This is chain-of-thought applied with structure — see our chain-of-thought guide for the underlying technique.

Think through the problem step by step inside
<thinking></thinking> tags. Then give your final
recommendation inside <answer></answer> tags.

Problem: {problem}

Exploit the long context window

Claude can hold very large documents in a single prompt, which makes it strong for reviewing contracts, synthesising research, or auditing a codebase in one pass. When you do this, tell Claude where to look: “Quote the exact clause before you summarise it” keeps long-context answers grounded in the source rather than paraphrased from memory.

Put durable rules in the system prompt

Claude follows system prompts closely, so it is the right place for persona, tone, and hard constraints. Read system prompt vs user prompt for the full pattern.

A starter set of Claude prompts

• Document Q&A: “Answer only from the <document> below. Quote the relevant text, then answer. If the answer is not present, say so.”
• Refactor reviewer: “Inside <thinking>, list risks of this refactor. Inside <answer>, give a go/no-go and the top 3 tests to add.”
• Tone-matched rewrite: “Rewrite the <draft> to match the voice in <example>. Keep all facts.”

For prompts that work across every provider, start from our best work prompts and add the tag structure above.