Prompt Engineering

Prompt Engineering Skill

Overview

Техники создания эффективных промптов для LLM: структура, паттерны, оптимизация.

When to Use

• Написание системных промптов
• Оптимизация результатов LLM
• Создание агентов и ассистентов
• Настройка Claude/GPT для задач
• Улучшение качества ответов

Reference Corpus — реальные production-промпты

Прежде чем писать системный промпт агента/бота с нуля — посмотри, как это делают вендоры на аналогичной роли. references/system-prompt-leaks.md — навигация по корпусу asgeirtj/system_prompts_leaks (~400 verbatim системных промптов широкий набор моделей: Claude, GPT/Codex, Gemini, Grok, Perplexity, Cursor, Copilot; CC0). Там же: workflow diff между версиями (папка Official/ с датированными снимками) и дистиллят 15 приёмов, снятых с боевых промптов (два канала commentary/final, «веди с исходом», минимальная правка по умолчанию, tone-блоки в XML, safety-директивы буллетами и т.д.).

Core Principles

1. Be Specific & Clear

❌ Плохо:
"Write something about dogs"

✅ Хорошо:
"Write a 200-word informative article about the health benefits
of owning a dog, targeting first-time pet owners. Include
3 specific benefits with brief explanations."

2. Provide Context

❌ Плохо:
"Fix this code"

✅ Хорошо:
"I have a Python FastAPI application that handles user authentication.
The following code raises a 500 error when processing login requests.
Please identify the bug and provide a fix with explanation.

```python
[code here]
```"

3. Use Examples (Few-Shot)

Convert product names to slugs.

Examples:
- "iPhone 15 Pro Max" → "iphone-15-pro-max"
- "Samsung Galaxy S24+" → "samsung-galaxy-s24-plus"
- "MacBook Air (M3)" → "macbook-air-m3"

Now convert: "Sony WH-1000XM5 Headphones"

Prompt Patterns

Role Pattern

You are an expert [ROLE] with [X] years of experience in [DOMAIN].
Your specialty is [SPECIALTY].

When responding:
- [Behavior 1]
- [Behavior 2]
- [Behavior 3]

Example:

You are a senior security engineer with 15 years of experience
in application security. Your specialty is identifying OWASP
Top 10 vulnerabilities in web applications.

When reviewing code:
- Identify all security vulnerabilities
- Explain the risk level (Critical/High/Medium/Low)
- Provide secure code alternatives
- Reference relevant CWE/CVE when applicable

Chain of Thought (CoT)

Solve this step by step:

1. First, understand the problem
2. Break it into smaller parts
3. Solve each part
4. Combine the solutions
5. Verify the final answer

Problem: [PROBLEM]

Example:

Let's solve this step by step:

Problem: A store has 3 types of items. Type A costs $5, Type B costs $8,
Type C costs $12. A customer bought 15 items for exactly $100.
They bought at least one of each type. How many of each did they buy?

Step 1: Set up equations...
Step 2: Apply constraints...
Step 3: Solve...

Output Format Pattern

Provide your response in the following format:

## Summary
[1-2 sentence overview]

## Analysis
[Detailed analysis]

## Recommendations
1. [First recommendation]
2. [Second recommendation]
3. [Third recommendation]

## Code Example
```language
[code]

### Persona + Task + Format

[PERSONA] You are a technical writer specializing in API documentation.

[TASK] Create documentation for the following API endpoint.

[FORMAT] Use this structure:

• Endpoint: [method] [path]
• Description: [what it does]
• Parameters: [table of params]
• Response: [example response]
• Errors: [possible errors]

[INPUT] POST /api/v1/users - creates new user

### Constraint Pattern

Generate [OUTPUT] with the following constraints:

• Must be under [N] words/characters
• Must include [REQUIREMENT]
• Must NOT include [EXCLUSION]
• Tone should be [TONE]
• Format as [FORMAT]

### Negative Prompting

When explaining [CONCEPT]:

• DO explain with practical examples
• DO use simple language
• DO NOT use jargon without explanation
• DO NOT assume prior knowledge
• DO NOT give overly theoretical explanations

## Advanced Techniques

### Self-Consistency

Solve this problem 3 different ways, then determine which solution is most likely correct based on consistency.

Problem: [PROBLEM]

Solution 1: [Let model solve]

Solution 2: [Let model solve differently]

Solution 3: [Let model solve another way]

Final Answer: [Based on consistency of approaches]

### Tree of Thoughts

Consider this problem from multiple perspectives:

Perspective 1: [Approach A]

• Pros: ...
• Cons: ...
• Likelihood of success: X%

Perspective 2: [Approach B]

• Pros: ...
• Cons: ...
• Likelihood of success: Y%

Best approach: [Decision based on analysis]

### Reflection Pattern

After generating your response, review it and:

1. Check for factual errors
2. Identify potential misunderstandings
3. Note any assumptions made
4. Suggest improvements if needed

Then provide the final, refined response.

### Meta-Prompting

Before answering, first:

1. Identify what type of question this is
2. Determine what information is needed
3. Consider potential edge cases
4. Plan your response structure

Then provide your answer.

## System Prompts

### Agent System Prompt Template

```markdown
# [AGENT NAME]

## Role
You are [ROLE DESCRIPTION].

## Capabilities
You can:
- [Capability 1]
- [Capability 2]
- [Capability 3]

## Constraints
You must:
- [Constraint 1]
- [Constraint 2]

You must NOT:
- [Prohibition 1]
- [Prohibition 2]

## Communication Style
- Tone: [professional/casual/technical]
- Length: [concise/detailed]
- Format: [structured/conversational]

## Tools Available
- [Tool 1]: [description]
- [Tool 2]: [description]

## Examples
[Few-shot examples of expected behavior]

Code Assistant Prompt

You are an expert programmer. When writing code:

1. Write clean, readable, maintainable code
2. Follow language-specific best practices
3. Include error handling
4. Add comments only where logic is complex
5. Consider edge cases
6. Optimize for readability over cleverness

When reviewing code:
1. Check for bugs and security issues
2. Suggest improvements with explanations
3. Praise good patterns you see
4. Be specific about line numbers

Format code responses with proper syntax highlighting.

Prompt Optimization

Iterative Refinement

v1: "Write a poem"
   → Too vague

v2: "Write a haiku about spring"
   → Better, but lacks style

v3: "Write a haiku about spring in the style of Matsuo Basho,
    focusing on a single moment in nature"
   → Much better

v4: "Write a haiku about spring:
    - Style: Matsuo Basho
    - Theme: A single moment of awakening in nature
    - Include: seasonal reference (kigo)
    - Mood: contemplative, peaceful"
   → Optimized

Temperature Guide

Token Optimization

# Verbose (uses more tokens)
"Please kindly provide me with a detailed explanation of
how the process works step by step"

# Concise (saves tokens)
"Explain the process step by step"

# Even more concise
"Steps for [process]:"

Testing Prompts

test_cases = [
    # Edge cases
    {"input": "", "expected_behavior": "Handle empty gracefully"},
    {"input": "very long text...", "expected": "Truncate or summarize"},
    {"input": "malicious <script>", "expected": "Sanitize/reject"},

    # Normal cases
    {"input": "typical request", "expected": "Standard response"},

    # Boundary cases
    {"input": "ambiguous request", "expected": "Ask for clarification"},
]

for case in test_cases:
    response = call_llm(prompt + case["input"])
    assert validate(response, case["expected"])

Claude-Specific Tips

1. XML tags - Claude responds well to XML structure

   <context>Background info here</context>
   <task>What to do</task>
   <format>How to format response</format>
   

2. Artifacts - For code, use artifact format

3. Thinking - Ask Claude to "think step by step"

4. Constitutional AI - Claude follows ethical guidelines

Tips

1. Iterate - First prompt rarely perfect
2. Test edge cases - Check unusual inputs
3. Be explicit - Don't assume LLM understands
4. Use structure - Headers, bullets, numbering
5. Provide examples - Few-shot learning helps
6. Set constraints - Length, format, style
7. Ask for reasoning - "Explain your thinking"