Context Engineering

Context Engineering

Overview

Context engineering is the practice of designing the entire context window—system prompts, examples, structure, instructions, and data—to maximize Claude's performance. This skill distills Anthropic's official prompt engineering documentation into actionable guidance.

When to Use

• Designing prompts for Claude API calls
• Writing system prompts for agents or assistants
• Structuring complex multi-part inputs
• Improving accuracy or consistency of outputs
• Adding examples to guide behavior
• Implementing reasoning patterns (chain of thought)

Technique Priority

Apply techniques in order of effectiveness. Not all tasks require all techniques.

Core Techniques

1. Be Clear and Direct

Think of Claude as a brilliant new employee who needs explicit instructions.

The Golden Rule: Show the prompt to a colleague with minimal context. If they're confused, Claude will be too.

Key Practices:

• Provide contextual information (what results are for, target audience, workflow position, success criteria)
• Be specific about desired output (format, length, style)
• Use numbered steps for sequential instructions

<example>
<poor>
Please remove all personally identifiable information from these messages.
</poor>
<good>
Your task is to anonymize customer feedback for our quarterly review.

Instructions:

1. Replace customer names with "CUSTOMER_[ID]"
2. Replace emails with "EMAIL_[ID]@example.com"
3. Redact phone numbers as "PHONE_[ID]"
4. Leave product names intact
5. Output only processed messages, separated by "---"

Data to process: {{FEEDBACK_DATA}} </good> </example>

2. Use Examples (Multishot Prompting)

Examples dramatically improve accuracy, consistency, and quality.

Best Practices:

• Include 3-5 diverse, relevant examples
• Cover edge cases and potential challenges
• Wrap examples in <example> tags (nest within <examples> if multiple)
• Vary examples enough to avoid unintended pattern matching

<example>
<prompt>
Our CS team needs to categorize feedback. Use categories: UI/UX, Performance, Feature Request, Integration, Pricing, Other. Rate sentiment (Positive/Neutral/Negative) and priority (High/Medium/Low).

<example>
Input: The new dashboard is a mess! It takes forever to load, and I can't find the export button. Fix this ASAP!
Category: UI/UX, Performance
Sentiment: Negative
Priority: High
</example>

Now analyze: {{FEEDBACK}} </prompt> </example>

3. Chain of Thought (CoT)

Encourage Claude to break down problems step-by-step for complex reasoning tasks.

When to Use:

• Math and calculations
• Multi-step analysis
• Logic problems
• Decisions with many factors

When to Avoid:

• Simple factual questions (adds latency without benefit)

Complexity Levels:

<example>
<basic>
Solve this problem. Think step-by-step.
</basic>
<structured>
Draft personalized donor emails.

Program info: {{PROGRAM_DETAILS}} Donor info: {{DONOR_DETAILS}}

Think before writing in <thinking> tags:

1. What messaging appeals to this donor given their history?
2. What program aspects match their interests?

Then write the email in <email> tags. </structured> </example>

4. XML Tags

Use XML tags to structure prompts with multiple components.

Benefits:

• Clarity: Separate instructions, examples, context, data
• Accuracy: Prevent Claude from mixing up components
• Flexibility: Easy to modify individual sections
• Parseability: Extract specific parts from outputs

Best Practices:

• Be consistent with tag names throughout prompts
• Nest tags for hierarchical content: <outer><inner></inner></outer>
• Reference tags in instructions: "Using the contract in <contract> tags..."
• Use meaningful names (<instructions>, <context>, <examples>, <data>)

<example>
<prompt>
Analyze this software licensing agreement for legal risks.

<context>
We're a multinational enterprise considering this for core infrastructure.
</context>

<agreement>
{{CONTRACT}}
</agreement>

<instructions>
1. Analyze: Indemnification, Limitation of liability, IP ownership
2. Note unusual or concerning terms
3. Compare to our standard: <standard_contract>{{STANDARD}}</standard_contract>
4. Summarize findings in <findings> tags
5. List recommendations in <recommendations> tags
</instructions>
</prompt>
</example>

See references/xml-tags.md for detailed patterns.

5. Role Prompting (System Prompts)

Use the system parameter to set Claude's role and dramatically improve domain performance.

Benefits:

• Enhanced accuracy in specialized domains
• Tailored communication style
• Improved focus on task requirements

Best Practices:

• Put role in system parameter, task in user turn
• Be specific: "data scientist specializing in customer insight for Fortune 500" vs "data scientist"
• Experiment with different roles for the same task

<example>
<basic>
system: "You are a helpful assistant."
</basic>
<enhanced>
system: "You are the General Counsel of a Fortune 500 tech company. You specialize in software licensing and data privacy regulations."
</enhanced>
</example>

6. Prefill Claude's Response

Guide outputs by prefilling the Assistant message.

Use Cases:

• Force specific output format (start with { for JSON)
• Skip preambles and explanations
• Maintain character in roleplay
• Ensure consistent structure

Constraints:

• Cannot end with trailing whitespace
• Not available with extended thinking mode

<example>
<json_output>
user: Extract name, price, color from: {{DESCRIPTION}}
assistant: {  <!-- prefill forces JSON output -->
</json_output>

<character_maintenance> user: What do you deduce about this shoe? assistant: [Sherlock Holmes] <!-- prefill maintains character --> </character_maintenance> </example>