customaize-agent:prompt-engineering（定制 Agent 提示工程）

Prompt Engineering Patterns

Advanced prompt engineering techniques to maximize LLM performance, reliability, and controllability.

Core Capabilities

1. Few-Shot Learning

Teach the model by showing examples instead of explaining rules. Include 2-5 input-output pairs that demonstrate the desired behavior. Use when you need consistent formatting, specific reasoning patterns, or handling of edge cases. More examples improve accuracy but consume tokens—balance based on task complexity.

Example:

Extract key information from support tickets:

Input: "My login doesn't work and I keep getting error 403"
Output: {"issue": "authentication", "error_code": "403", "priority": "high"}

Input: "Feature request: add dark mode to settings"
Output: {"issue": "feature_request", "error_code": null, "priority": "low"}

Now process: "Can't upload files larger than 10MB, getting timeout"

2. Chain-of-Thought Prompting

Request step-by-step reasoning before the final answer. Add "Let's think step by step" (zero-shot) or include example reasoning traces (few-shot). Use for complex problems requiring multi-step logic, mathematical reasoning, or when you need to verify the model's thought process. Improves accuracy on analytical tasks by 30-50%.

Example:

Analyze this bug report and determine root cause.

Think step by step:
1. What is the expected behavior?
2. What is the actual behavior?
3. What changed recently that could cause this?
4. What components are involved?
5. What is the most likely root cause?

Bug: "Users can't save drafts after the cache update deployed yesterday"

3. Prompt Optimization

Systematically improve prompts through testing and refinement. Start simple, measure performance (accuracy, consistency, token usage), then iterate. Test on diverse inputs including edge cases. Use A/B testing to compare variations. Critical for production prompts where consistency and cost matter.

Example:

Version 1 (Simple): "Summarize this article"
→ Result: Inconsistent length, misses key points

Version 2 (Add constraints): "Summarize in 3 bullet points"
→ Result: Better structure, but still misses nuance

Version 3 (Add reasoning): "Identify the 3 main findings, then summarize each"
→ Result: Consistent, accurate, captures key information

4. Template Systems

Build reusable prompt structures with variables, conditional sections, and modular components. Use for multi-turn conversations, role-based interactions, or when the same pattern applies to different inputs. Reduces duplication and ensures consistency across similar tasks.

Example:

# Reusable code review template
template = """
Review this {language} code for {focus_area}.

Code:
{code_block}

Provide feedback on:
{checklist}
"""

# Usage
prompt = template.format(
    language="Python",
    focus_area="security vulnerabilities",
    code_block=user_code,
    checklist="1. SQL injection\n2. XSS risks\n3. Authentication"
)

5. System Prompt Design

Set global behavior and constraints that persist across the conversation. Define the model's role, expertise level, output format, and safety guidelines. Use system prompts for stable instructions that shouldn't change turn-to-turn, freeing up user message tokens for variable content.

Example:

System: You are a senior backend engineer specializing in API design.

Rules:
- Always consider scalability and performance
- Suggest RESTful patterns by default
- Flag security concerns immediately
- Provide code examples in Python
- Use early return pattern

Format responses as:
1. Analysis
2. Recommendation
3. Code example
4. Trade-offs

Key Patterns

Progressive Disclosure

Start with simple prompts, add complexity only when needed:

1. Level 1: Direct instruction

- "Summarize this article"

1. Level 2: Add constraints

- "Summarize this article in 3 bullet points, focusing on key findings"

1. Level 3: Add reasoning

- "Read this article, identify the main findings, then summarize in 3 bullet points"

1. Level 4: Add examples

- Include 2-3 example summaries with input-output pairs

Instruction Hierarchy

[System Context] → [Task Instruction] → [Examples] → [Input Data] → [Output Format]

Error Recovery

Build prompts that gracefully handle failures:

• Include fallback instructions
• Request confidence scores
• Ask for alternative interpretations when uncertain
• Specify how to indicate missing information

Best Practices

1. Be Specific: Vague prompts produce inconsistent results
2. Show, Don't Tell: Examples are more effective than descriptions
3. Test Extensively: Evaluate on diverse, representative inputs
4. Iterate Rapidly: Small changes can have large impacts
5. Monitor Performance: Track metrics in production
6. Version Control: Treat prompts as code with proper versioning
7. Document Intent: Explain why prompts are structured as they are

Common Pitfalls

• Over-engineering: Starting with complex prompts before trying simple ones
• Example pollution: Using examples that don't match the target task
• Context overflow: Exceeding token limits with excessive examples
• Ambiguous instructions: Leaving room for multiple interpretations
• Ignoring edge cases: Not testing on unusual or boundary inputs

Integration Patterns

With RAG Systems

# Combine retrieved context with prompt engineering
prompt = f"""Given the following context:
{retrieved_context}

{few_shot_examples}

Question: {user_question}

Provide a detailed answer based solely on the context above. If the context doesn't contain enough information, explicitly state what's missing."""

With Validation

# Add self-verification step
prompt = f"""{main_task_prompt}

After generating your response, verify it meets these criteria:
1. Answers the question directly
2. Uses only information from provided context
3. Cites specific sources
4. Acknowledges any uncertainty

If verification fails, revise your response."""

Performance Optimization

Token Efficiency

• Remove redundant words and phrases
• Use abbreviations consistently after first definition
• Consolidate similar instructions
• Move stable content to system prompts

Latency Reduction

• Minimize prompt length without sacrificing quality
• Use streaming for long-form outputs
• Cache common prompt prefixes
• Batch similar requests when possible

Agent Prompting Best Practices

Based on Anthropic's official best practices for agent prompting.

Core principles

Context Window

The “context window” refers to the entirety of the amount of text a language model can look back on and reference when generating new text plus the new text it generates. This is different from the large corpus of data the language model was trained on, and instead represents a “working memory” for the model. A larger context window allows the model to understand and respond to more complex and lengthy prompts, while a smaller context window may limit the model’s ability to handle longer prompts or maintain coherence over extended conversations.

• Progressive token accumulation: As the conversation advances through turns, each user message and assistant response accumulates within the context window. Previous turns are preserved completely.
• Linear growth pattern: The context usage grows linearly with each turn, with previous turns preserved completely.
• 200K token capacity: The total available context window (200,000 tokens) represents the maximum capacity for storing conversation history and generating new output from Claude.
• Input-output flow: Each turn consists of:

- Input phase: Contains all previous conversation history plus the current user message - Output phase: Generates a text response that becomes part of a future input

Concise is key

The context window is a public good. Your prompt, command, skill shares the context window with everything else Claude needs to know, including:

• The system prompt
• Conversation history
• Other commands, skills, hooks, metadata
• Your actual request

Default assumption: Claude is already very smart

Only add context Claude doesn't already have. Challenge each piece of information:

• "Does Claude really need this explanation?"
• "Can I assume Claude knows this?"
• "Does this paragraph justify its token cost?"

Good example: Concise (approximately 50 tokens):

## Extract PDF text

Use pdfplumber for text extraction:

import pdfplumber

with pdfplumber.open("file.pdf") as pdf: text = pdf.pages[0].extract_text()

Bad example: Too verbose (approximately 150 tokens):

## Extract PDF text

PDF (Portable Document Format) files are a common file format that contains
text, images, and other content. To extract text from a PDF, you'll need to
use a library. There are many libraries available for PDF processing, but we
recommend pdfplumber because it's easy to use and handles most cases well.
First, you'll need to install it using pip. Then you can use the code below...

The concise version assumes Claude knows what PDFs are and how libraries work.

Set appropriate degrees of freedom

Match the level of specificity to the task's fragility and variability.

High freedom (text-based instructions):

Use when:

• Multiple approaches are valid
• Decisions depend on context
• Heuristics guide the approach

Example:

## Code review process

1. Analyze the code structure and organization
2. Check for potential bugs or edge cases
3. Suggest improvements for readability and maintainability
4. Verify adherence to project conventions

Medium freedom (pseudocode or scripts with parameters):

Use when:

• A preferred pattern exists
• Some variation is acceptable
• Configuration affects behavior

Example:

## Generate report

Use this template and customize as needed:

def generate_report(data, format="markdown", include_charts=True): # Process data # Generate output in specified format # Optionally include visualizations

Low freedom (specific scripts, few or no parameters):

Use when:

• Operations are fragile and error-prone
• Consistency is critical
• A specific sequence must be followed

Example:

## Database migration

Run exactly this script:

python scripts/migrate.py --verify --backup

Do not modify the command or add additional flags.

Analogy: Think of Claude as a robot exploring a path:

• Narrow bridge with cliffs on both sides: There's only one safe way forward. Provide specific guardrails and exact instructions (low freedom). Example: database migrations that must run in exact sequence.
• Open field with no hazards: Many paths lead to success. Give general direction and trust Claude to find the best route (high freedom). Example: code reviews where context determines the best approach.

Persuasion Principles for Agent Communication

Usefull for writing prompts, including but not limited to: commands, hooks, skills for Claude Code, or prompts for sub agents or any other LLM interaction.

Overview

LLMs respond to the same persuasion principles as humans. Understanding this psychology helps you design more effective skills - not to manipulate, but to ensure critical practices are followed even under pressure.

Research foundation: Meincke et al. (2025) tested 7 persuasion principles with N=28,000 AI conversations. Persuasion techniques more than doubled compliance rates (33% → 72%, p <.001).

The Seven Principles

1. Authority

What it is: Deference to expertise, credentials, or official sources.

How it works in prompts:

• Imperative language: "YOU MUST", "Never", "Always"
• Non-negotiable framing: "No exceptions"
• Eliminates decision fatigue and rationalization

When to use:

• Discipline-enforcing skills (TDD, verification requirements)
• Safety-critical practices
• Established best practices

Example:

✅ Write code before test? Delete it. Start over. No exceptions.
❌ Consider writing tests first when feasible.

2. Commitment

What it is: Consistency with prior actions, statements, or public declarations.

How it works in prompts:

• Require announcements: "Announce skill usage"
• Force explicit choices: "Choose A, B, or C"
• Use tracking: TodoWrite for checklists

When to use:

• Ensuring skills are actually followed
• Multi-step processes
• Accountability mechanisms

Example:

✅ When you find a skill, you MUST announce: "I'm using [Skill Name]"
❌ Consider letting your partner know which skill you're using.

3. Scarcity

What it is: Urgency from time limits or limited availability.

How it works in prompts:

• Time-bound requirements: "Before proceeding"
• Sequential dependencies: "Immediately after X"
• Prevents procrastination

When to use:

• Immediate verification requirements
• Time-sensitive workflows
• Preventing "I'll do it later"

Example:

✅ After completing a task, IMMEDIATELY request code review before proceeding.
❌ You can review code when convenient.

4. Social Proof

What it is: Conformity to what others do or what's considered normal.

How it works in prompts:

• Universal patterns: "Every time", "Always"
• Failure modes: "X without Y = failure"
• Establishes norms

When to use:

• Documenting universal practices
• Warning about common failures
• Reinforcing standards

Example:

✅ Checklists without TodoWrite tracking = steps get skipped. Every time.
❌ Some people find TodoWrite helpful for checklists.

5. Unity

What it is: Shared identity, "we-ness", in-group belonging.

How it works in prompts:

• Collaborative language: "our codebase", "we're colleagues"
• Shared goals: "we both want quality"

When to use:

• Collaborative workflows
• Establishing team culture
• Non-hierarchical practices

Example:

✅ We're colleagues working together. I need your honest technical judgment.
❌ You should probably tell me if I'm wrong.

6. Reciprocity

What it is: Obligation to return benefits received.

How it works:

• Use sparingly - can feel manipulative
• Rarely needed in prompts

When to avoid:

• Almost always (other principles more effective)

7. Liking

What it is: Preference for cooperating with those we like.

How it works:

• DON'T USE for compliance
• Conflicts with honest feedback culture
• Creates sycophancy

When to avoid:

• Always for discipline enforcement

Principle Combinations by Prompt Type

Why This Works: The Psychology

Bright-line rules reduce rationalization:

• "YOU MUST" removes decision fatigue
• Absolute language eliminates "is this an exception?" questions
• Explicit anti-rationalization counters close specific loopholes

Implementation intentions create automatic behavior:

• Clear triggers + required actions = automatic execution
• "When X, do Y" more effective than "generally do Y"
• Reduces cognitive load on compliance

LLMs are parahuman:

• Trained on human text containing these patterns
• Authority language precedes compliance in training data
• Commitment sequences (statement → action) frequently modeled
• Social proof patterns (everyone does X) establish norms

Ethical Use

Legitimate:

• Ensuring critical practices are followed
• Creating effective documentation
• Preventing predictable failures

Illegitimate:

• Manipulating for personal gain
• Creating false urgency
• Guilt-based compliance

The test: Would this technique serve the user's genuine interests if they fully understood it?

Quick Reference

When designing a prompt, ask:

1. What type is it? (Discipline vs. guidance vs. reference)
2. What behavior am I trying to change?
3. Which principle(s) apply? (Usually authority + commitment for discipline)
4. Am I combining too many? (Don't use all seven)
5. Is this ethical? (Serves user's genuine interests?)