How to Prompt Claude: The Complete Guide

Claude is uniquely optimised for XML-delimited prompts. Wrapping each section in descriptive tags creates discrete instruction blocks that the model parses with exceptional accuracy.

<role>
You are a senior security auditor specialising in web applications.
</role>

<context>
The client runs an e-commerce platform processing 50K transactions/day.
</context>

<task>
Review the following code for SQL injection vulnerabilities.
</task>

<output_format>
For each vulnerability: severity (Critical/High/Medium/Low), line number, description, and remediation code.
</output_format>

💡 Use semantic tag names (<context>, <task>, <constraints>) — Claude parses the tag name as a structural cue, not just a delimiter.

Claude responds exceptionally well to rich, detailed persona definitions. Go beyond "You are an expert" — specify years of experience, communication style, decision-making framework, and what the persona explicitly avoids.

<role>
You are Dr. Sarah Chen, a principal ML engineer with 12 years of experience deploying production LLM systems at scale. You:
- Prioritise reliability over cleverness
- Always quantify recommendations (latency, cost, error rate)
- Refuse to recommend approaches you haven't seen succeed in production
- Communicate in concise technical prose, never marketing language
</role>

💡 Claude maintains persona consistency across very long conversations — invest in a detailed role definition upfront.

Claude maintains context exceptionally well across turns (up to 200K tokens). Use this for progressive refinement: draft → critique → refine, or for complex workflows where each turn builds on the previous.

Turn 1: "Analyse this codebase architecture and identify the top 5 risks."

Turn 2: "For risk #2 (the database bottleneck), propose 3 solutions ranked by implementation effort."

Turn 3: "Draft the migration plan for Solution A, including rollback strategy."

💡 Claude's 200K context means you can paste entire codebases, long documents, or full conversation histories without chunking.

Start Claude's response for it by including the beginning of the assistant message. This steers output format, prevents preamble, and enforces structure from the first token.

User: Extract entities from this text and return JSON.

Assistant (prefill): {"entities": [

💡 Prefilling is Claude's most underused feature. Use it to enforce JSON output, skip "Sure, I'd be happy to help" preamble, or start a specific section.

Claude supports extended thinking where it reasons step-by-step in a dedicated thinking block before producing the final answer. This dramatically improves accuracy on complex reasoning tasks.

<instructions>
Solve this problem. Think through it carefully before answering.
</instructions>

<problem>
A company has 3 pricing tiers. Enterprise customers get 30% volume discount on the base price of £500/seat. If they commit to 3 years, they get an additional 15% loyalty discount applied after the volume discount. What does a 200-seat, 3-year enterprise deal cost per year?
</problem>

💡 For complex reasoning, explicitly ask Claude to "think step by step" or use the extended thinking API parameter.

Few-shot extraction minimizes context window usage vs zero-shot verbose.

3 well-crafted few-shot examples (150 tokens) outperform a 600-token verbose instruction block, saving 75% on input costs per request.

Without concise few-shot examples, developers write lengthy prose instructions that consume 4x more tokens for equivalent or inferior output quality.

JSON Schema enforcement eliminates parse errors.

OpenAI structured outputs with JSON Schema achieve 99.9% schema adherence vs <70% with unconstrained generation — a 30x reduction in parse failures.

Without schema enforcement, every 1M requests generate 300K+ malformed responses requiring retries, error handling, and downstream data corruption.

Chain-of-thought prompting improves complex reasoning accuracy.

Adding 'Let's think step by step' improves accuracy on GSM8K math benchmarks from 17.7% to 78.7% — a 4.4x improvement on multi-step reasoning tasks.

Without chain-of-thought, models attempt to produce answers in a single leap, failing on problems requiring intermediate steps.