Mastering Claude Prompt Engineering: 9 Proven Strategies to Boost LLM Performance

Anthropic Prompt Optimization Strategies

Anthropic officially recommends seven key strategies to get the most out of Claude and other advanced large language models:

Be clear and direct

Use examples (multishot)

Let Claude think (chain of thought)

Use XML tags

Prefill Claude’s response

Chain complex prompts

Long‑context tips

Key Takeaways from Personal Experience

Structure prompts with XML tags to clearly separate context, instructions, and examples.

Trigger Claude’s thinking and planning by asking it to reason step‑by‑step before answering.

Provide contextual information such as task purpose, audience, workflow position, and success criteria.

Specify exactly what Claude should do , e.g., output only code or JSON.

Order instructions using numbered or bulleted lists to ensure correct execution.

XML‑Based Prompt Structure

When prompts contain multiple types of information (context, examples, instructions), wrap each part in custom XML tags. Example:

<example>…</example>
<examples>…</examples>

Maintain consistency and nesting, e.g., <outer><inner>…</inner></outer>.

Multishot Prompting

Multishot prompting is a concrete implementation of few‑shot learning for LLMs; the two are not opposites.

Provide well‑designed examples to improve accuracy, consistency, and quality, especially for structured output.

<example>…</example>
<example>…</example>

Chain‑of‑Thought (CoT)

For complex tasks, ask Claude to decompose the problem step‑by‑step, which yields more precise and detailed responses.

Prefilling Claude’s Response

Start the assistant message with a partial answer so Claude continues from there, allowing you to skip long introductions, enforce specific output formats (e.g., JSON), or maintain role consistency.

{
  "model": "claude-sonnet-4-20250514",
  "max_tokens": 1024,
  "messages": [
    {"role": "user", "content": "请帮我写一封感谢信…"},
    {"role": "assistant", "content": "Dear [Name],

我希望这封邮件能找到你。"}
  ]
}

Prompt Chaining

Break a complex task into smaller sub‑tasks, feed the output of one step as the input to the next, and iterate for optimization.

1. Analyze the existing code and identify refactoring points.
2. Design a new modular structure with function signatures.
3. Implement each function step‑by‑step.

Example XML‑structured output for a refactoring analysis:

<refactoring_points>
  <point>
    <description>函数职责单一原则违反</description>
    <reason>函数承担过多职责，难以测试</reason>
    <improvement>拆分为 validate、communicate、record 等独立函数</improvement>
  </point>
</refactoring_points>

Extended Thinking

Enable Claude’s extended thinking mode via the API to obtain a transparent reasoning trace before the final answer.

response = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=16000,
    thinking={"type": "enabled", "budget_tokens": 10000},
    messages=[{"role": "user", "content": "Are there an infinite number of prime numbers such that n mod 4 == 3?"}]
)

The response includes a thinking block followed by the final text block.

High‑Level Instructions

Claude often performs better when given a concise high‑level goal rather than overly prescriptive step‑by‑step directions, allowing it to apply its own creativity.

Practical Coding Example

Ask Claude to write a Python function with test cases, then have it run the tests in its extended thinking phase.

编写一个计算数字阶乘的函数。
在完成之前，请使用以下测试用例验证你的解决方案：
- n=0
- n=1
- n=5
- n=10
并修复你发现的任何问题。

Reference Links

Prompt Engineering Overview

Be Clear and Direct

Multishot Prompting

Chain of Thought

Use XML Tags

Prefill Claude’s Response

Chain Prompts

Long Context Tips

Extended Thinking