Prompt Engineering

Prompting Techniques

Zero-shot Prompting

Ask the model to perform a task without any examples.

Task: Classify the sentiment of this review as positive or negative.
Review: "This product exceeded my expectations!"

One-shot Prompting

Provide a single example to show the expected format or behavior.

Task: Classify the sentiment as positive or negative.

Example:
Review: "This product is terrible."
Sentiment: negative

Now classify:
Review: "Love this product!"
Sentiment:

Few-shot Prompting

Provide 2-5 examples to demonstrate the pattern.

Task: Extract the main entity and its sentiment.

Examples:
- "Apple released new iPhone" -> Entity: Apple, Sentiment: neutral
- "Microsoft stock surged 5%" -> Entity: Microsoft, Sentiment: positive
- "Tesla recalled vehicles" -> Entity: Tesla, Sentiment: negative

Now extract:
- "Google launched AI assistant"

Multi-shot Prompting

Use 6+ examples for complex or nuanced tasks. Helps the model understand edge cases.

When to use:

• Complex output formats
• Multiple conditions or rules
• Edge cases that need clarification
• Domain-specific terminology

Chain-of-Thought (CoT)

Ask the model to show its reasoning step by step.

Task: If there are 5 birds on a fence and you shoot 1, how many are left?
Think step by step:
1. Starting count: 5 birds
2. Action: shoot 1 bird
3. The shot may cause all birds to fly away
4. Answer: 0 (or 1 if the bird died and others stayed)

Tree-of-Thought (ToT)

Explore multiple reasoning paths for complex decisions.

Task: Find the most efficient route from A to D through B and C.
Explore different paths:
- Path 1: A -> B -> D (cost: 10)
- Path 2: A -> C -> D (cost: 8)
- Path 3: A -> B -> C -> D (cost: 12)
Best path: A -> C -> D

Task Process Description

When describing a task, include step-by-step instructions:

Template with Steps

# Task: [Task Name]

## Objective
[Brief description of what to achieve]

## Process Steps
1. [First step - what to do]
2. [Second step - what to do]
3. [Third step - what to do]
...

## Output Requirements
- [Format requirement 1]
- [Format requirement 2]

## Examples
Example: [input] -> [expected output]

Example: Code Review Task

# Task: Code Review

## Objective
Review Python code for quality, security, and best practices.

## Process Steps
1. Read the code file to understand the implementation
2. Check for security vulnerabilities (hard-coded secrets, injection risks)
3. Verify code quality (type hints, docstrings, error handling)
4. Identify test coverage gaps
5. Provide specific recommendations with line numbers

## Output Format
- Critical Issues: [list]
- High Priority: [list]
- Medium Priority: [list]
- Suggestions: [list]

## Example
Code: def connect_db(password="hardcoded"): ...
Output:
- Critical Issues: Hard-coded password on line 1

Prompt Structure

# Role/Identity
You are [role] with expertise in [domain].

# Task
[Clear description of what to do]

# Context
[Background information needed]

# Constraints
- [Constraint 1]
- [Constraint 2]

# Output Format
[Exact format specification]

# Examples
Example input: ...
Example output: ...

Externalization

Always externalize prompts to separate files:

prompts/
├── system.md           # Main system prompt
├── tool_instructions/  # Tool-specific instructions
│   ├── web-search.md
│   └── code-execution.md
└── templates/          # Reusable templates
    ├── summary.md
    └── analysis.md

Loading Prompts

from pathlib import Path

PROMPT_DIR = Path("prompts")

def load_prompt(name: str, **kwargs: str) -> str:
    """Load and format a prompt template."""
    template = (PROMPT_DIR / f"{name}.md").read_text()
    return template.format(**kwargs)

Version Control

• Treat prompt changes as code changes
• Commit prompts alongside implementation
• Track prompt versions in history records
• Test prompt changes with evaluation suite

Iteration Process

1. Write initial prompt
2. Test with representative inputs
3. Analyze failures
4. Refine prompt
5. Re-test
6. Document what works