Claude Prompt Engineering Best Practices: System Prompts, Few-Shot Examples & Chain-of-Thought Techniques

Maximize Claude’s Response Quality with Proven Prompt Engineering Techniques

Getting the best results from Claude requires more than just asking questions. Strategic prompt engineering—through well-designed system prompts, carefully placed few-shot examples, and chain-of-thought reasoning—can dramatically improve output accuracy, consistency, and relevance. This guide covers practical, workflow-oriented techniques you can implement immediately.

1. System Prompt Design: Setting the Foundation

The system prompt establishes Claude’s persona, constraints, and output expectations before any user interaction begins. A well-structured system prompt is the single most impactful lever for response quality.

Anatomy of an Effective System Prompt

import anthropic

• Start with a one-sentence summary.
• Follow with code blocks.
• End with potential pitfalls or edge cases.""", messages=[ {“role”: “user”, “content”: “How should I implement rate limiting for my API?”} ] ) print(response.content[0].text)

System Prompt Structure Checklist

• Role definition — Who is Claude in this context?- Behavioral constraints — What should Claude always or never do?- Output format specification — Structure, length, and style expectations.- Domain boundaries — What topics are in or out of scope?- Fallback behavior — How to handle ambiguity or missing information.

2. Few-Shot Example Placement: Teaching by Demonstration

Few-shot prompting gives Claude concrete input-output pairs so it can pattern-match your expectations. Placement and quality of examples matter significantly.

Basic Few-Shot Pattern

response = client.messages.create( model=“claude-sonnet-4-20250514”, max_tokens=512, system=“You extract structured data from unstructured product reviews.”, messages=[ {“role”: “user”, “content”: “Review: ‘The battery lasts forever but the screen is too dim outdoors.’”}, {“role”: “assistant”, “content”: ’{“sentiment”: “mixed”, “pros”: [“battery life”], “cons”: [“screen brightness outdoors”], “score”: 3.5}’}, {“role”: “user”, “content”: “Review: ‘Absolutely terrible. Broke after two days and customer support ghosted me.’”}, {“role”: “assistant”, “content”: ’{“sentiment”: “negative”, “pros”: [], “cons”: [“durability”, “customer support”], “score”: 1.0}’}, {“role”: “user”, “content”: “Review: ‘Best purchase this year. Fast shipping, great build quality, and the app integration is seamless.’”} ] ) print(response.content[0].text)

Few-Shot Placement Rules

Chain-of-thought prompting instructs Claude to show its reasoning process before arriving at a conclusion. This is critical for math, logic, multi-step analysis, and decision-making tasks.

Explicit CoT with XML Tags

response = client.messages.create( model=“claude-sonnet-4-20250514”, max_tokens=2048, system="""You are a financial analyst. When answering questions:

The user will NOT see the block, so put your complete final response in .""", messages=[ {“role”: “user”, “content”: “A company has revenue of $2.4M, COGS of $1.1M, and operating expenses of $800K. What is the operating margin, and is it healthy for a SaaS startup?”} ] ) print(response.content[0].text)

Extended Thinking (Built-in CoT)

Claude models support a native extended thinking feature via the API, which allocates a dedicated reasoning budget before generating the response. response = client.messages.create( model="claude-sonnet-4-20250514", max_tokens=8000, thinking={ "type": "enabled", "budget_tokens": 5000 }, messages=[ {"role": "user", "content": "Design a database schema for a multi-tenant SaaS application with row-level security."} ] )

for block in response.content: if block.type == “thinking”: print(“[Reasoning]”, block.thinking) elif block.type == “text”: print(“[Answer]”, block.text)

4. Installation & Setup

Get started with the Anthropic Python SDK: # Install the SDK pip install anthropic

Set your API key as an environment variable

python -c “import anthropic; print(anthropic.version)“

Or use the API directly via cURL: curl https://api.anthropic.com/v1/messages
-H “x-api-key: YOUR_API_KEY”
-H “content-type: application/json”
-H “anthropic-version: 2023-06-01”
-d ’{ “model”: “claude-sonnet-4-20250514”, “max_tokens”: 1024, “system”: “You are a helpful coding assistant.”, “messages”: [{“role”: “user”, “content”: “Explain async/await in Python.”}] }‘

5. Pro Tips for Power Users

• Use XML tags for structure — Claude responds exceptionally well to XML-delimited sections like , , and <output_format> within prompts.- Prefill the assistant turn — Start Claude’s response by providing an opening in the assistant message to steer format (e.g., {“role”: “assistant”, “content”: ”{”} forces JSON output).- Separate data from instructions — Place long documents or data inside clearly labeled XML tags so Claude doesn’t confuse content with instructions.- Temperature tuning — Use temperature=0 for deterministic tasks (data extraction, classification) and temperature=0.7–1.0 for creative writing or brainstorming.- Batch API for scale — For high-volume prompt workflows, use the Message Batches API to process thousands of prompts at 50% reduced cost.- Cache system prompts — Use prompt caching with the cache_control parameter to reduce latency and cost when reusing large system prompts.

6. Troubleshooting Common Issues

What is the ideal length for a Claude system prompt?

There is no hard limit, but aim for 200–800 words for most use cases. Claude can handle system prompts exceeding 10,000 tokens effectively, especially with prompt caching enabled. The key is clarity and structure—use sections, numbered rules, and XML tags rather than writing dense paragraphs. Longer system prompts work well when they contain reference material, but keep behavioral instructions concise and front-loaded.

How many few-shot examples should I include for best results?

For most tasks, 2–4 examples strike the best balance between quality and token efficiency. One example establishes the pattern, two confirm it, and three to four handle edge cases. For highly nuanced tasks like sentiment analysis with custom scales, go up to 5–6 examples. Beyond that, returns diminish and you consume tokens that could be used for the actual response. Always include at least one edge case or negative example.

When should I use extended thinking versus manual chain-of-thought prompting?

Use extended thinking (the thinking parameter) when you want Claude to reason internally without exposing the reasoning to end users—ideal for production applications. Use manual CoT with XML tags like when you need to inspect, debug, or log the reasoning process during development. Extended thinking is also more effective for extremely complex tasks because it allocates dedicated compute to reasoning before the response generation begins.