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.

Priority	Technique	Best For
1	Be clear and direct	All tasks
2	Use examples (multishot)	Format consistency, complex patterns
3	Chain of thought	Math, logic, analysis, complex reasoning
4	XML tags	Multi-part prompts, structured I/O
5	Role prompting	Domain expertise, tone adjustment
6	Prefill response	Output format control, character consistency
7	Chain prompts	Multi-step workflows, error isolation
8	Long context tips	Documents >20K tokens
9	Extended thinking	Complex STEM, constraint optimization

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

Please remove all personally identifiable information from these messages. 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}}

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

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). 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

Now analyze: {{FEEDBACK}}

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:

Level	Approach	Example
Basic	"Think step-by-step"	Quick, less guided
Guided	Outline specific steps	More control over reasoning
Structured	Use <thinking> and <answer> tags	Easy to parse, separates reasoning from output

Solve this problem. Think step-by-step. Draft personalized donor emails.

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

Think before writing in tags:

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

Then write the email in tags.

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>)

Analyze this software licensing agreement for legal risks. We're a multinational enterprise considering this for core infrastructure. {{CONTRACT}} 1. Analyze: Indemnification, Limitation of liability, IP ownership 2. Note unusual or concerning terms 3. Compare to our standard: {{STANDARD}} 4. Summarize findings in tags 5. List recommendations in tags

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

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

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

user: Extract name, price, color from: {{DESCRIPTION}} assistant: {

<character_maintenance> user: What do you deduce about this shoe? assistant: [Sherlock Holmes] </character_maintenance>

7. Chain Complex Prompts

Break complex tasks into sequential subtasks for better accuracy.

When to Chain:

• Multi-step analysis or synthesis
• Content creation pipelines
• Tasks requiring self-correction
• Complex transformations

Benefits:

• Each subtask gets full attention
• Easier to debug specific steps
• Can parallelize independent subtasks

Patterns:

• Research → Outline → Draft → Edit → Format
• Extract → Transform → Analyze → Visualize
• Generate → Review → Refine → Re-review (self-correction)

Prompt 1: Analyze contract for risks → {{ANALYSIS}} Prompt 2: Draft email based on {{ANALYSIS}} Prompt 3: Review email for tone and clarity → {{FEEDBACK}} Prompt 4: Revise email based on {{FEEDBACK}}

8. Long Context Tips

For prompts with substantial data (20K+ tokens):

Key Practices:

1. Put data at the top: Place long documents above queries/instructions (up to 30% quality improvement)

2. Structure with XML: Wrap documents with metadata

<documents>
  <document index="1">
    <source>annual_report.pdf</source>
    <document_content>{{CONTENT}}</document_content>
  </document>
</documents>

3. Ground in quotes: Ask Claude to quote relevant sections before answering

See references/long-context.md for detailed patterns.

9. Extended Thinking

For complex problems requiring deep reasoning:

Best Practices:

• Start with minimum budget (1024 tokens), increase as needed
• Use general instructions first ("think thoroughly"), not prescriptive steps
• Ask Claude to verify work with test cases
• Use batch processing for >32K thinking tokens

Best Use Cases:

• Complex STEM problems
• Constraint optimization (multiple competing requirements)
• Problems requiring structured frameworks

See references/extended-thinking.md for detailed patterns.

Quick Reference

Output Format Control

Goal	Technique
JSON output	Prefill with {
Specific structure	Provide example in <example> tags
No preamble	Prefill or explicit instruction
Consistent format	Multishot examples

Improving Accuracy

Problem	Solution
Misses instructions	Number steps, be explicit
Inconsistent format	Add examples
Wrong reasoning	Add CoT with structured output
Misses context	Add role prompting
Drops steps	Chain into separate prompts

Refactoring Existing Prompts

When optimizing or compressing an existing prompt, apply these checks after every structural change:

Pitfall	Check
Stale cross-references	After renaming or renumbering steps, search for ALL references to old labels (jump targets, "see Step X", resume points) and update them
Over-abstraction	If the model needs exact values to execute (specific keys, field names, command arguments), keep them literal even if they look repetitive -- a generic placeholder the model cannot expand is worse than duplication
Lost preconditions	When merging or removing steps, verify that any precondition checks or guards in the removed step are preserved elsewhere
Silent behavior changes	Diff the before/after and confirm every deleted line is either redundant or relocated, not dropped

Common Tag Names

Tag	Purpose
<instructions>	Task directives
<context>	Background information
<example> / <examples>	Few-shot demonstrations
<data> / <document>	Input content
<thinking>	Chain of thought reasoning
<answer> / <output>	Final response
<constraints>	Limitations or requirements

Resources

For detailed guidance on specific techniques:

• references/xml-tags.md - Comprehensive XML structuring patterns
• references/chain-of-thought.md - CoT implementation details
• references/long-context.md - Working with large documents
• references/extended-thinking.md - Extended thinking mode