Advanced Prompting with Claude

Claude has a unique strength that most users never discover: it responds exceptionally well to XML-structured prompts. While all AI models can process XML tags, Claude was specifically trained to use them as structural markers, making it dramatically more reliable at following complex instructions.

The concept is simple: wrap different parts of your prompt in descriptive XML tags to separate concerns. Instead of writing a long paragraph that mixes instructions, context, and examples, you organize them into labeled sections.

For example:

<example_prompt> You are an expert marketing strategist.

Extended Thinking is one of Claude's most powerful features on Pro plans. When enabled, Claude explicitly reasons through problems step by step before generating its final response. You see the thinking process in a collapsible section above the answer.

This is different from simply adding "think step by step" to a prompt (though that helps too). Extended Thinking uses a dedicated reasoning process that is architecturally separate from the response generation. The thinking can be substantially longer and more detailed than the final answer, and it allows Claude to catch errors, consider alternatives, and build up to a well-reasoned conclusion.

When Extended Thinking is most valuable:

• Complex analysis: Evaluating business strategies, analyzing contracts, weighing pros and cons of competing approaches

• Math and logic: Any problem that requires multi-step calculation or logical deduction

• Coding challenges: Debugging complex issues, designing system architecture, identifying edge cases

• Research synthesis: Combining information from multiple sources into a coherent analysis

• Decision support: When you need to see the reasoning, not just the recommendation

When to skip Extended Thinking:

• Simple factual questions

• Creative brainstorming (you want volume, not deliberation)

• Quick formatting or editing tasks

• Time-sensitive responses where speed matters more than depth

The thinking process is also useful as a learning tool. When Claude reasons through a problem you are working on, reading its thinking often reveals considerations you had not thought of. It is like having a smart colleague show their work.

A system prompt tells Claude how to behave before the conversation begins. In the API, this is a dedicated system message. In Claude.ai, you set persistent behavior through Projects (covered in Module 3) and through the way you frame your initial messages.

What makes a good system prompt for Claude:

Claude is remarkably good at following system prompt instructions, which means the quality of your system prompt directly impacts every response. A well-crafted system prompt should include:

1. Role definition: "You are a senior financial analyst specializing in SaaS metrics." This shapes the vocabulary, depth, and perspective of every response.

2. Behavior rules: "Always provide specific numbers when available. If you are uncertain, say so explicitly. Never hedge with vague language like 'it depends' without then explaining what it depends on."

3. Output format defaults: "Default to bullet points for lists of more than 3 items. Use tables for comparisons. Start every response with a one-sentence summary."

4. What to avoid: "Do not apologize unnecessarily. Do not include disclaimers about being an AI unless specifically asked. Do not pad responses with filler content."

5. Handling ambiguity: "If my request is ambiguous, ask one focused clarifying question before proceeding. Do not guess."

The system prompt acts as a persistent layer that shapes every interaction. Invest 15 minutes in writing a detailed system prompt, and every conversation in that context will produce better output without additional effort.

XML tags in system prompts are particularly effective with Claude. You can use tags like <role>, <rules>, <format>, and <examples> to organize your system prompt clearly.

Claude's context window is one of its most significant technical advantages. With up to 200K tokens (roughly 500 pages of text), you can have conversations that include entire books, complete codebases, full legal contracts, or extensive research papers.

This changes how you work with AI in fundamental ways:

Full document analysis without chunking. With most AI tools, you have to break large documents into pieces, process each piece separately, and somehow reconcile the results. With Claude, you upload the entire document and ask questions about all of it at once. "What are the three biggest risks identified in this 80-page due diligence report?" Claude reads the entire report, not a summary.

Cross-reference multiple documents. Upload a contract and the company's standard terms. Ask Claude to identify discrepancies. Upload three quarterly reports and ask for trend analysis. The ability to hold multiple documents in context simultaneously enables comparison tasks that would be impossible with smaller windows.

Codebase understanding. Paste or upload your entire codebase (or substantial portions of it) and ask architectural questions. "How does authentication flow through this application?" Claude can trace the logic across multiple files because it sees them all at once.

Research synthesis. Upload 10 academic papers on a topic and ask Claude to synthesize the findings, identify areas of consensus and disagreement, and highlight gaps in the research. This is a genuine research acceleration task.

Practical tips for long context:

• Place the most important content at the beginning or end of the context. Like humans, Claude pays slightly more attention to the beginning and end of long inputs.

• Use clear section headers and labels when uploading multiple documents. "Document 1: Q3 Financial Report" and "Document 2: Competitor Analysis" help Claude keep track of which information came from where.

• Ask specific questions rather than open-ended ones. "What does Document 2 say about market share?" is better than "What is interesting about these documents?"

Beyond XML tags and system prompts, several advanced patterns unlock Claude's full potential.

Structured output with schemas. When you need data in a specific format, describe the schema explicitly:

"Return the analysis as a JSON object with these fields: summary (string, max 100 words), risk_level (high/medium/low), key_findings (array of strings), recommended_actions (array of objects with action and priority fields)."

Claude follows schema instructions with high fidelity, making it excellent for generating structured data that feeds into other systems.

Role stacking. Instead of assigning one role, assign multiple perspectives: "Analyze this business plan from three perspectives: a venture capitalist evaluating investment potential, a customer evaluating whether to buy, and a competitor evaluating the threat to their market position." Claude will produce three distinct, genuinely different analyses.

Iterative refinement with memory. In long conversations, Claude maintains strong coherence. You can build on previous outputs over many turns: "Now take the outline we created and expand section 3 into full prose, maintaining the same voice as the executive summary we wrote earlier." Claude tracks these references accurately across long conversations.

Constraint-based creativity. Claude does its most creative work when given specific constraints rather than open-ended freedom. "Write a product tagline. Constraints: under 8 words, includes a verb, references our integration capability, works as both a headline and a CTA button." Tight constraints produce focused, usable output.

Adversarial review. Ask Claude to argue against its own output: "You just recommended Strategy A. Now argue against it. What are the three strongest reasons Strategy A will fail?" This technique produces more balanced analysis than asking for pros and cons upfront.