05: Prompt Engineering Basics

Home›Series›Claude for PHP Developers›Chapter 05

Chapter 05: Prompt Engineering Basics

Overview

Prompt engineering is the art and science of crafting effective instructions for AI models. The quality of Claude's responses is directly proportional to the quality of your prompts. A well-engineered prompt can mean the difference between generic, unhelpful outputs and precise, actionable results.

This chapter provides comprehensive coverage of prompt engineering fundamentals, including clear instruction writing, few-shot learning, chain-of-thought reasoning, role-playing techniques, structured output formats, and practical PHP implementation patterns. You'll learn to write prompts that consistently produce high-quality results.

What You'll Learn:

Core principles of effective prompting
Few-shot learning with positive and negative examples
Chain-of-thought reasoning techniques
Role-playing and persona assignment
Structured output formatting (JSON, XML, CSV)
Prompt templates and reusability
Prompt length optimization and token management
Iterative prompt refinement methodologies
Systematic prompt testing and evaluation
Common pitfalls and how to avoid them

Prerequisites

Before starting, ensure you have:

✓ Completed Chapters 01-04
✓ PHP 8.4+ with good understanding
✓ Anthropic API key configured
✓ Experience with Claude API from previous chapters

Estimated Time: ~45-55 minutes

What You'll Build

By the end of this chapter, you will have created:

A FewShotClassifier class for sentiment analysis and text classification
A ChainOfThoughtSolver class for complex problem-solving with reasoning
A PersonaBuilder class for creating dynamic AI personas and roles
A PromptTemplate class for reusable prompt templates with variable substitution
A PromptLibrary class for managing a collection of pre-built prompts
A PromptLengthManager class for optimizing prompt size and token usage
A PromptRefiner class for iterative prompt improvement
A PromptTester class for systematic prompt testing and evaluation
A PromptVersionTracker class for tracking prompt changes over time
Multiple example scripts demonstrating prompt engineering techniques
Understanding of few-shot learning (including negative examples), chain-of-thought reasoning, and role-playing
Knowledge of structured output formats (JSON, XML, CSV, Markdown)
Skills in prompt optimization, iterative refinement, and testing methodologies
Best practices for writing effective prompts that produce consistent results

Objectives

By the end of this chapter, you will be able to:

Write clear, specific, and context-rich prompts that produce high-quality results
Implement few-shot learning techniques including positive and negative examples
Apply chain-of-thought reasoning for complex problem-solving tasks
Create dynamic personas and role-based prompts for specialized use cases
Generate structured outputs in JSON, XML, CSV, and Markdown formats
Build reusable prompt templates and libraries for consistent prompting
Optimize prompt length and manage token usage effectively
Refine prompts iteratively using systematic testing and analysis
Test prompts systematically with proper metrics and A/B testing
Identify and avoid common prompt engineering pitfalls
Track prompt versions and measure improvements over time

Principles of Effective Prompts

1. Be Clear and Specific

Vague prompts produce vague results. Be explicit about what you want.

php

<?php
# filename: examples/01-clarity.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

// ✗ Bad: Vague prompt
$vagueResponse = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 1024,
    'messages' => [[
        'role' => 'user',
        'content' => 'Tell me about PHP.'
    ]]
]);

echo "Vague prompt result:\n";
echo $vagueResponse->content[0]->text . "\n\n";

// ✓ Good: Specific prompt
$specificResponse = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 1024,
    'messages' => [[
        'role' => 'user',
        'content' => 'Explain the benefits of PHP 8.4\'s typed class constants in 3 bullet points. Include one code example.'
    ]]
]);

echo "Specific prompt result:\n";
echo $specificResponse->content[0]->text . "\n\n";

2. Provide Context

Give Claude the information it needs to understand your request.

php

<?php
# filename: examples/02-context.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

$userCode = <<<'PHP'
class UserRepository {
    public function find($id) {
        return DB::table('users')->where('id', $id)->first();
    }
}
PHP;

// ✗ Bad: No context
$noContextResponse = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 1024,
    'messages' => [[
        'role' => 'user',
        'content' => "Review this code:\n\n{$userCode}"
    ]]
]);

// ✓ Good: With context
$withContextResponse = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 1024,
    'messages' => [[
        'role' => 'user',
        'content' => "You are reviewing a Laravel 11 application that handles sensitive user data. Review this repository class for security issues, type safety, and error handling:\n\n{$userCode}"
    ]]
]);

echo "With context:\n";
echo $withContextResponse->content[0]->text . "\n";

3. Specify Output Format

Tell Claude exactly how you want the response formatted.

php

<?php
# filename: examples/03-output-format.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

$response = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 1024,
    'messages' => [[
        'role' => 'user',
        'content' => 'List 5 PHP design patterns. Format as JSON with fields: name, category, use_case, difficulty. Return only valid JSON.'
    ]]
]);

echo $response->content[0]->text . "\n";

// Parse the JSON
$text = $response->content[0]->text;
if (preg_match('/```json\s*(\{.*?\}|\[.*?\])\s*```/s', $text, $matches)) {
    $json = $matches[1];
} else {
    $json = $text;
}

$data = json_decode($json, true);
print_r($data);

4. Use Delimiters for Clarity

Separate instructions from data using clear delimiters.

php

<?php
# filename: examples/04-delimiters.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

$emailContent = "Hi, I need help with my order #12345. It hasn't arrived yet. Please refund ASAP!";

$response = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 512,
    'messages' => [[
        'role' => 'user',
        'content' => <<<PROMPT
Analyze the customer email below and extract:
1. Order number
2. Issue category (delivery, refund, product, etc.)
3. Urgency level (low, medium, high)
4. Sentiment (positive, neutral, negative)

Return as JSON.

--- EMAIL STARTS ---
{$emailContent}
--- EMAIL ENDS ---
PROMPT
    ]]
]);

echo $response->content[0]->text . "\n";

Few-Shot Learning

Few-shot learning involves providing examples to guide Claude's responses.

Zero-Shot (No Examples)

php

<?php
# filename: examples/05-zero-shot.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

// No examples provided
$response = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 512,
    'messages' => [[
        'role' => 'user',
        'content' => 'Classify this text as positive, negative, or neutral: "The product is okay, nothing special."'
    ]]
]);

echo "Zero-shot classification:\n";
echo $response->content[0]->text . "\n\n";

One-Shot (One Example)

php

<?php
# filename: examples/06-one-shot.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

$response = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 512,
    'messages' => [[
        'role' => 'user',
        'content' => <<<PROMPT
Classify customer feedback sentiment.

Example:
Input: "This product exceeded my expectations! Absolutely love it."
Output: positive

Now classify this:
Input: "The product is okay, nothing special."
Output:
PROMPT
    ]]
]);

echo "One-shot classification:\n";
echo $response->content[0]->text . "\n\n";

Few-Shot (Multiple Examples)

php

<?php
# filename: examples/07-few-shot.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

$response = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 512,
    'messages' => [[
        'role' => 'user',
        'content' => <<<PROMPT
Classify customer feedback sentiment as: positive, negative, or neutral.

Examples:

Input: "This product exceeded my expectations! Absolutely love it."
Output: positive

Input: "Terrible quality, broke after one day. Very disappointed."
Output: negative

Input: "The product is okay, nothing special."
Output: neutral

Input: "Fast shipping and good packaging."
Output: positive

Now classify this:
Input: "Not what I expected, but it works fine."
Output:
PROMPT
    ]]
]);

echo "Few-shot classification:\n";
echo $response->content[0]->text . "\n\n";

Few-Shot with Structured Output

php

<?php
# filename: src/Services/FewShotClassifier.php
declare(strict_types=1);

namespace App\Services;

use Anthropic\Anthropic;

class FewShotClassifier
{
    public function __construct(
        private readonly Anthropic $client
    ) {}

    public function classify(string $text, array $examples, array $categories): string
    {
        $examplesText = $this->formatExamples($examples);
        $categoriesText = implode(', ', $categories);

        $prompt = <<<PROMPT
Classify the text into one of these categories: {$categoriesText}

Examples:
{$examplesText}

Now classify this:
Input: {$text}
Output:
PROMPT;

        $response = $this->client->messages()->create([
            'model' => 'claude-sonnet-4-20250514',
            'max_tokens' => 100,
            'temperature' => 0.0, // Deterministic for classification
            'messages' => [[
                'role' => 'user',
                'content' => $prompt
            ]]
        ]);

        return trim($response->content[0]->text);
    }

    private function formatExamples(array $examples): string
    {
        $formatted = [];

        foreach ($examples as $example) {
            $formatted[] = "Input: {$example['input']}\nOutput: {$example['output']}";
        }

        return implode("\n\n", $formatted);
    }
}

Usage:

php

<?php
# filename: examples/08-few-shot-classifier.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use App\Services\FewShotClassifier;
use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

$classifier = new FewShotClassifier($client);

$examples = [
    ['input' => 'This Laravel package is amazing!', 'output' => 'positive'],
    ['input' => 'Terrible documentation, bugs everywhere.', 'output' => 'negative'],
    ['input' => 'It works as expected.', 'output' => 'neutral'],
];

$result = $classifier->classify(
    text: 'Great performance improvements in this version!',
    examples: $examples,
    categories: ['positive', 'negative', 'neutral']
);

echo "Classification: {$result}\n";

Negative Examples (What NOT to Do)

Sometimes showing Claude what not to do is as effective as showing what to do. Negative examples help establish boundaries and clarify edge cases.

php

<?php
# filename: examples/08b-negative-examples.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

// Classification with negative examples
$response = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 512,
    'messages' => [[
        'role' => 'user',
        'content' => <<<PROMPT
Classify customer feedback as: positive, negative, or neutral.

Positive examples:
Input: "This product exceeded my expectations! Absolutely love it."
Output: positive

Input: "Fast shipping and excellent customer service."
Output: positive

Negative examples (what NOT to classify as positive):
Input: "It's okay, nothing special."
Output: neutral (NOT positive - this is neutral feedback)

Input: "The product works but I expected more."
Output: neutral (NOT positive - mixed sentiment)

Now classify this:
Input: "Not what I expected, but it works fine."
Output:
PROMPT
    ]]
]);

echo "Classification with negative examples:\n";
echo $response->content[0]->text . "\n\n";

When to use negative examples:

Edge case clarification: When boundaries between categories are unclear
Error prevention: To prevent common misclassifications
Constraint definition: To establish what responses are unacceptable
Quality control: To filter out low-quality or inappropriate outputs

php

<?php
# filename: examples/08c-negative-examples-code-review.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

$code = <<<'PHP'
function getUser($id) {
    return DB::query("SELECT * FROM users WHERE id = " . $id);
}
PHP;

$response = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 1024,
    'messages' => [[
        'role' => 'user',
        'content' => <<<PROMPT
Review this PHP code for security issues.

Good security practices:
- Use prepared statements
- Validate and sanitize input
- Use parameterized queries

Bad practices (what NOT to do):
- String concatenation in SQL queries (SQL injection risk)
- Direct user input in database queries
- No input validation

Code to review:
{$code}

Identify security issues and explain why they're problematic.
PROMPT
    ]]
]);

echo $response->content[0]->text . "\n";

Chain-of-Thought Reasoning

Chain-of-thought prompting encourages Claude to show its reasoning process.

Basic Chain-of-Thought

php

<?php
# filename: examples/09-chain-of-thought.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

// Without chain-of-thought
$direct = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 512,
    'messages' => [[
        'role' => 'user',
        'content' => 'Should I use a singleton pattern for my database connection in PHP?'
    ]]
]);

echo "Direct answer:\n";
echo $direct->content[0]->text . "\n\n";

// With chain-of-thought
$cot = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 1024,
    'messages' => [[
        'role' => 'user',
        'content' => 'Should I use a singleton pattern for my database connection in PHP? Think through this step by step, considering: 1) Performance implications, 2) Testing challenges, 3) Modern alternatives like dependency injection.'
    ]]
]);

echo "Chain-of-thought answer:\n";
echo $cot->content[0]->text . "\n\n";

Explicit Step-by-Step Instructions

php

<?php
# filename: examples/10-explicit-steps.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

$code = <<<'PHP'
function processOrder($order) {
    $total = 0;
    foreach ($order['items'] as $item) {
        $total += $item['price'] * $item['quantity'];
    }
    if ($order['discount']) {
        $total = $total - ($total * $order['discount']);
    }
    return $total;
}
PHP;

$response = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 2048,
    'messages' => [[
        'role' => 'user',
        'content' => <<<PROMPT
Refactor this PHP function following these steps:

1. Identify issues (type safety, error handling, naming)
2. List specific improvements needed
3. Provide refactored code with PHPDoc
4. Explain the changes made

Code to refactor:
{$code}
PROMPT
    ]]
]);

echo $response->content[0]->text . "\n";

Self-Consistency Chain-of-Thought

php

<?php
# filename: src/Services/ChainOfThoughtSolver.php
declare(strict_types=1);

namespace App\Services;

use Anthropic\Anthropic;

class ChainOfThoughtSolver
{
    public function __construct(
        private readonly Anthropic $client
    ) {}

    public function solve(string $problem, int $attempts = 3): array
    {
        $solutions = [];

        for ($i = 0; $i < $attempts; $i++) {
            $response = $this->client->messages()->create([
                'model' => 'claude-sonnet-4-20250514',
                'max_tokens' => 2048,
                'temperature' => 0.7, // Some variation
                'messages' => [[
                    'role' => 'user',
                    'content' => <<<PROMPT
{$problem}

Think through this step by step:
1. Analyze the problem
2. Consider different approaches
3. Evaluate trade-offs
4. Recommend solution
5. Provide implementation

Format your response with clear sections.
PROMPT
                ]]
            ]);

            $solutions[] = $response->content[0]->text;
        }

        return [
            'solutions' => $solutions,
            'consensus' => $this->findConsensus($solutions)
        ];
    }

    private function findConsensus(array $solutions): string
    {
        // Ask Claude to find consensus among solutions
        $combined = implode("\n\n--- SOLUTION ---\n\n", $solutions);

        $response = $this->client->messages()->create([
            'model' => 'claude-sonnet-4-20250514',
            'max_tokens' => 1024,
            'messages' => [[
                'role' => 'user',
                'content' => <<<PROMPT
I have multiple solutions to a problem. Identify the consensus or best approach:

{$combined}

Provide a concise summary of the best solution.
PROMPT
            ]]
        ]);

        return $response->content[0]->text;
    }
}

Role-Playing and Personas

Assign Claude a specific role or expertise to improve response quality.

Basic Role Assignment

php

<?php
# filename: examples/11-roles.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

// Different roles produce different perspectives
$roles = [
    'senior_developer' => 'You are a senior PHP developer with 15 years of experience. Focus on maintainability and best practices.',
    'security_expert' => 'You are a security expert specializing in web applications. Focus on vulnerabilities and security implications.',
    'performance_engineer' => 'You are a performance optimization engineer. Focus on speed, efficiency, and scalability.',
];

$code = <<<'PHP'
function getUser($id) {
    return DB::query("SELECT * FROM users WHERE id = " . $id);
}
PHP;

foreach ($roles as $roleName => $rolePrompt) {
    $response = $client->messages()->create([
        'model' => 'claude-sonnet-4-20250514',
        'max_tokens' => 1024,
        'system' => $rolePrompt,
        'messages' => [[
            'role' => 'user',
            'content' => "Review this code:\n\n{$code}"
        ]]
    ]);

    echo "=== {$roleName} ===\n";
    echo $response->content[0]->text . "\n\n";
}

Multi-Attribute Personas

php

<?php
# filename: examples/12-personas.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

$persona = <<<PERSONA
You are an expert PHP architect with these attributes:
- 10+ years experience with Laravel
- Specialized in e-commerce platforms
- Strong advocate for SOLID principles
- Prefers composition over inheritance
- Values type safety and static analysis
- Writes comprehensive tests

When reviewing code:
1. Focus on architecture and design patterns
2. Suggest improvements for maintainability
3. Provide specific, actionable feedback
4. Include code examples
PERSONA;

$code = <<<'PHP'
class OrderProcessor {
    public function process($order) {
        // Validate order
        if (!$order) return false;

        // Calculate total
        $total = 0;
        foreach ($order->items as $item) {
            $total += $item->price;
        }

        // Save to database
        DB::save($order);

        // Send email
        Mail::send($order->email, 'Order confirmed');

        return true;
    }
}
PHP;

$response = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 3000,
    'system' => $persona,
    'messages' => [[
        'role' => 'user',
        'content' => "Review this OrderProcessor class and suggest architectural improvements:\n\n{$code}"
    ]]
]);

echo $response->content[0]->text . "\n";

Persona Builder Pattern

php

<?php
# filename: src/Services/PersonaBuilder.php
declare(strict_types=1);

namespace App\Services;

class PersonaBuilder
{
    private string $role = '';
    private int $experience = 0;
    private array $specializations = [];
    private array $principles = [];
    private array $guidelines = [];

    public function role(string $role): self
    {
        $this->role = $role;
        return $this;
    }

    public function experience(int $years): self
    {
        $this->experience = $years;
        return $this;
    }

    public function specialization(string $specialization): self
    {
        $this->specializations[] = $specialization;
        return $this;
    }

    public function principle(string $principle): self
    {
        $this->principles[] = $principle;
        return $this;
    }

    public function guideline(string $guideline): self
    {
        $this->guidelines[] = $guideline;
        return $this;
    }

    public function build(): string
    {
        $parts = [];

        if ($this->role) {
            $experienceText = $this->experience > 0
                ? " with {$this->experience}+ years of experience"
                : '';

            $parts[] = "You are a {$this->role}{$experienceText}.";
        }

        if (!empty($this->specializations)) {
            $parts[] = "\nSpecializations:";
            foreach ($this->specializations as $spec) {
                $parts[] = "- {$spec}";
            }
        }

        if (!empty($this->principles)) {
            $parts[] = "\nCore principles you follow:";
            foreach ($this->principles as $principle) {
                $parts[] = "- {$principle}";
            }
        }

        if (!empty($this->guidelines)) {
            $parts[] = "\nWhen responding:";
            foreach ($this->guidelines as $i => $guideline) {
                $parts[] = ($i + 1) . ". {$guideline}";
            }
        }

        return implode("\n", $parts);
    }
}

Usage:

php

<?php
# filename: examples/13-persona-builder.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use App\Services\PersonaBuilder;
use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

$persona = (new PersonaBuilder())
    ->role('Laravel Security Consultant')
    ->experience(12)
    ->specialization('OWASP Top 10 vulnerabilities')
    ->specialization('Authentication and authorization')
    ->specialization('Input validation and sanitization')
    ->principle('Security first, always')
    ->principle('Defense in depth')
    ->principle('Least privilege principle')
    ->guideline('Identify specific vulnerabilities')
    ->guideline('Explain the risk level')
    ->guideline('Provide secure code examples')
    ->guideline('Reference OWASP guidelines')
    ->build();

echo "Generated Persona:\n{$persona}\n\n";

$code = <<<'PHP'
Route::get('/user/{id}', function($id) {
    $user = DB::select("SELECT * FROM users WHERE id = " . $id);
    return view('profile', ['user' => $user]);
});
PHP;

$response = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 2048,
    'system' => $persona,
    'messages' => [[
        'role' => 'user',
        'content' => "Review this Laravel route for security issues:\n\n{$code}"
    ]]
]);

echo "Security Review:\n";
echo $response->content[0]->text . "\n";

Structured Output Formats

JSON Output

php

<?php
# filename: examples/14-json-output.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

$response = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 1024,
    'messages' => [[
        'role' => 'user',
        'content' => <<<PROMPT
Extract information from this text and return as JSON:

"John Doe, senior developer at Acme Corp, can be reached at john@acme.com or 555-1234. He specializes in PHP and Laravel."

Required JSON structure:
{
    "name": "string",
    "title": "string",
    "company": "string",
    "email": "string",
    "phone": "string",
    "skills": ["array", "of", "strings"]
}

Return only valid JSON, no explanation.
PROMPT
    ]]
]);

$text = $response->content[0]->text;

// Extract JSON
if (preg_match('/```json\s*(\{.*?\})\s*```/s', $text, $matches)) {
    $jsonText = $matches[1];
} else {
    $jsonText = $text;
}

$data = json_decode($jsonText, true);

echo "Extracted data:\n";
print_r($data);

XML Output

php

<?php
# filename: examples/15-xml-output.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

$response = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 1024,
    'messages' => [[
        'role' => 'user',
        'content' => <<<PROMPT
Generate PHPUnit test configuration XML for a Laravel project with these requirements:
- Test suite: Unit tests in tests/Unit
- Test suite: Feature tests in tests/Feature
- Coverage: HTML report in coverage/
- Stop on failure: yes
- Colors: enabled

Return only valid XML, no explanation.
PROMPT
    ]]
]);

$xml = $response->content[0]->text;

// Extract XML from markdown if wrapped
if (preg_match('/```xml\s*(.*?)\s*```/s', $xml, $matches)) {
    $xml = $matches[1];
}

echo $xml . "\n";

// Validate XML
$doc = new DOMDocument();
if ($doc->loadXML($xml)) {
    echo "\n✓ Valid XML generated\n";
}

CSV Output

php

<?php
# filename: examples/16-csv-output.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

$response = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 1024,
    'messages' => [[
        'role' => 'user',
        'content' => <<<PROMPT
Generate a CSV file with 5 sample PHP developers with these columns:
name, email, years_experience, primary_framework, hourly_rate

Include header row. Return only CSV, no explanation.
PROMPT
    ]]
]);

$csv = $response->content[0]->text;

// Extract CSV from markdown if wrapped
if (preg_match('/```csv\s*(.*?)\s*```/s', $csv, $matches)) {
    $csv = $matches[1];
} elseif (preg_match('/```\s*(.*?)\s*```/s', $csv, $matches)) {
    $csv = $matches[1];
}

echo $csv . "\n";

// Parse CSV
$lines = str_getcsv($csv, "\n");
$data = array_map('str_getcsv', $lines);

echo "\nParsed data:\n";
print_r($data);

Markdown Output

php

<?php
# filename: examples/17-markdown-output.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

$response = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 2048,
    'messages' => [[
        'role' => 'user',
        'content' => <<<PROMPT
Create technical documentation for a Laravel middleware that handles API rate limiting.

Format as Markdown with:
- Title (h1)
- Overview section
- Installation instructions
- Configuration section with code example
- Usage examples
- Best practices
PROMPT
    ]]
]);

$markdown = $response->content[0]->text;

echo $markdown . "\n";

// Optionally save to file
file_put_contents('middleware-docs.md', $markdown);
echo "\n✓ Saved to middleware-docs.md\n";

Prompt Templates

Template Engine

php

<?php
# filename: src/Services/PromptTemplate.php
declare(strict_types=1);

namespace App\Services;

class PromptTemplate
{
    public function __construct(
        private readonly string $template
    ) {}

    public function render(array $variables): string
    {
        $result = $this->template;

        foreach ($variables as $key => $value) {
            $placeholder = "{{" . $key . "}}";
            $result = str_replace($placeholder, $value, $result);
        }

        return $result;
    }

    public static function codeReview(): self
    {
        return new self(<<<'TEMPLATE'
You are a {{role}} with {{experience}} years of experience.

Review the following {{language}} code for:
1. {{aspect1}}
2. {{aspect2}}
3. {{aspect3}}

Code to review:
{{code}}

Provide:
- List of issues found
- Severity of each issue (critical, high, medium, low)
- Specific recommendations
- Refactored code if needed
TEMPLATE
        );
    }

    public static function dataExtraction(): self
    {
        return new self(<<<'TEMPLATE'
Extract the following information from the text below:
{{fields}}

Return as JSON with this structure:
{{json_structure}}

Text to analyze:
--- START ---
{{text}}
--- END ---

Return only valid JSON, no explanation.
TEMPLATE
        );
    }

    public static function documentation(): self
    {
        return new self(<<<'TEMPLATE'
Generate {{doc_type}} documentation for this {{code_type}}:

{{code}}

Include:
{{requirements}}

Format: {{format}}
TEMPLATE
        );
    }
}

Usage:

php

<?php
# filename: examples/18-templates.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use App\Services\PromptTemplate;
use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

// Code review template
$template = PromptTemplate::codeReview();

$prompt = $template->render([
    'role' => 'Senior Laravel Developer',
    'experience' => '10',
    'language' => 'PHP',
    'aspect1' => 'Security vulnerabilities',
    'aspect2' => 'Performance issues',
    'aspect3' => 'Code maintainability',
    'code' => 'function getUser($id) { return DB::query("SELECT * FROM users WHERE id = " . $id); }'
]);

echo "Generated prompt:\n{$prompt}\n\n";

$response = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 2048,
    'messages' => [['role' => 'user', 'content' => $prompt]]
]);

echo "Response:\n{$response->content[0]->text}\n";

Prompt Library

php

<?php
# filename: src/Services/PromptLibrary.php
declare(strict_types=1);

namespace App\Services;

class PromptLibrary
{
    private const PROMPTS = [
        'code_review' => [
            'template' => 'Review this {{language}} code for {{focus}}:\n\n{{code}}',
            'defaults' => ['language' => 'PHP', 'focus' => 'best practices']
        ],
        'bug_fix' => [
            'template' => 'Debug this {{language}} code. The error is: {{error}}\n\nCode:\n{{code}}',
            'defaults' => ['language' => 'PHP']
        ],
        'optimization' => [
            'template' => 'Optimize this {{language}} code for {{metric}}:\n\n{{code}}',
            'defaults' => ['language' => 'PHP', 'metric' => 'performance']
        ],
        'documentation' => [
            'template' => 'Generate {{doc_type}} for this {{language}} {{code_type}}:\n\n{{code}}',
            'defaults' => ['doc_type' => 'PHPDoc', 'language' => 'PHP', 'code_type' => 'function']
        ],
        'test_generation' => [
            'template' => 'Generate {{test_framework}} tests for this {{language}} {{code_type}}:\n\n{{code}}',
            'defaults' => ['test_framework' => 'PHPUnit', 'language' => 'PHP', 'code_type' => 'class']
        ],
    ];

    public static function get(string $name, array $variables = []): string
    {
        if (!isset(self::PROMPTS[$name])) {
            throw new \InvalidArgumentException("Prompt template '{$name}' not found");
        }

        $prompt = self::PROMPTS[$name];
        $merged = array_merge($prompt['defaults'] ?? [], $variables);

        return self::renderTemplate($prompt['template'], $merged);
    }

    private static function renderTemplate(string $template, array $variables): string
    {
        foreach ($variables as $key => $value) {
            $template = str_replace("{{" . $key . "}}", $value, $template);
        }

        return $template;
    }

    public static function list(): array
    {
        return array_keys(self::PROMPTS);
    }
}

Usage:

php

<?php
# filename: examples/19-prompt-library.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use App\Services\PromptLibrary;
use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

// List available prompts
echo "Available prompts:\n";
foreach (PromptLibrary::list() as $name) {
    echo "- {$name}\n";
}
echo "\n";

// Use a prompt
$code = <<<'PHP'
function calculateTotal($items) {
    $total = 0;
    foreach ($items as $item) {
        $total += $item['price'];
    }
    return $total;
}
PHP;

$prompt = PromptLibrary::get('test_generation', [
    'code' => $code,
    'test_framework' => 'PHPUnit',
    'code_type' => 'function'
]);

echo "Generated prompt:\n{$prompt}\n\n";

$response = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 2048,
    'messages' => [['role' => 'user', 'content' => $prompt]]
]);

echo "Generated tests:\n{$response->content[0]->text}\n";

Prompt Length and Token Management

Understanding prompt length and token limits is crucial for effective prompt engineering. Claude has a context window limit, and managing your prompt size ensures you have room for responses and can control costs.

Understanding Token Limits

Claude models have different context window sizes:

Claude Sonnet 4: Up to 200,000 tokens (~150,000 words)
Claude Opus 4: Up to 200,000 tokens
Claude Haiku 3: Up to 200,000 tokens

Important considerations:

Both input (prompt) and output (response) count toward token usage
Longer prompts = higher costs
Very long prompts may reduce response quality
Reserve tokens for the response (max_tokens)

Estimating Prompt Length

php

<?php
# filename: examples/20-token-estimation.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

// Simple token estimation (rough: 1 token ≈ 4 characters)
function estimateTokens(string $text): int
{
    return (int) ceil(strlen($text) / 4);
}

$prompt = <<<PROMPT
You are a PHP code reviewer. Review this code for security issues:

function getUser(\$id) {
    return DB::query("SELECT * FROM users WHERE id = " . \$id);
}
PROMPT;

$estimatedTokens = estimateTokens($prompt);
echo "Estimated prompt tokens: {$estimatedTokens}\n";
echo "Recommended max_tokens: " . max(512, 200000 - $estimatedTokens - 1000) . "\n";

Optimizing Prompt Length

Strategy 1: Remove Redundancy

php

<?php
# filename: examples/21-prompt-optimization.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

// ✗ Bad: Verbose and repetitive
$verbosePrompt = <<<PROMPT
You are a PHP developer. You need to review PHP code.
Please review the following PHP code.
Look for security issues in the PHP code.
Review the code carefully for any security vulnerabilities.
PROMPT;

// ✓ Good: Concise and clear
$concisePrompt = <<<PROMPT
Review this PHP code for security vulnerabilities:

function getUser(\$id) {
    return DB::query("SELECT * FROM users WHERE id = " . \$id);
}
PROMPT;

echo "Verbose: " . strlen($verbosePrompt) . " chars\n";
echo "Concise: " . strlen($concisePrompt) . " chars\n";
echo "Saved: " . (strlen($verbosePrompt) - strlen($concisePrompt)) . " characters\n";

Strategy 2: Use Abbreviations and Shorthand

php

<?php
# filename: examples/22-prompt-compression.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

// For classification tasks, use compact format
$compactPrompt = <<<PROMPT
Classify sentiment: pos/neg/neu

Examples:
"Great!" → pos
"Terrible" → neg
"Okay" → neu

Classify: "Not bad, could be better"
PROMPT;

// Claude understands abbreviations in context
$response = $client->messages()->create([
    'model' => 'claude-sonnet-4-20250514',
    'max_tokens' => 50,
    'messages' => [['role' => 'user', 'content' => $compactPrompt]]
]);

echo $response->content[0]->text . "\n";

Strategy 3: Prioritize Important Information

php

<?php
# filename: examples/23-prompt-prioritization.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

// Put critical instructions first
$optimizedPrompt = <<<PROMPT
Extract: name, email, phone from text below. Return JSON only.

Text:
John Doe, john@example.com, 555-1234
PROMPT;

// Less important context can be shortened or omitted if space is tight

Managing Long Contexts

When working with long documents or codebases:

php

<?php
# filename: src/Services/PromptLengthManager.php
declare(strict_types=1);

namespace App\Services;

class PromptLengthManager
{
    private const MAX_PROMPT_TOKENS = 150000; // Reserve space for response
    private const TOKENS_PER_CHAR = 4; // Rough estimate

    public function truncateIfNeeded(string $prompt, int $maxTokens = self::MAX_PROMPT_TOKENS): string
    {
        $estimatedTokens = $this->estimateTokens($prompt);

        if ($estimatedTokens <= $maxTokens) {
            return $prompt;
        }

        // Truncate intelligently (preserve structure)
        $maxChars = ($maxTokens * self::TOKENS_PER_CHAR) - 100; // Safety margin

        // Try to preserve important parts (instructions, examples)
        if (preg_match('/^(.*?)(Examples?:.*?)(Text:|Code:.*)$/s', $prompt, $matches)) {
            $instructions = $matches[1];
            $examples = $matches[2];
            $data = $matches[3];

            $availableForData = $maxChars - strlen($instructions) - strlen($examples);
            $truncatedData = substr($data, 0, $availableForData) . "\n\n[... truncated ...]";

            return $instructions . $examples . $truncatedData;
        }

        // Fallback: simple truncation
        return substr($prompt, 0, $maxChars) . "\n\n[... truncated ...]";
    }

    private function estimateTokens(string $text): int
    {
        return (int) ceil(strlen($text) / self::TOKENS_PER_CHAR);
    }

    public function getRecommendedMaxTokens(string $prompt): int
    {
        $promptTokens = $this->estimateTokens($prompt);
        $available = self::MAX_PROMPT_TOKENS - $promptTokens - 1000; // Safety margin
        return max(512, min(8192, $available)); // Reasonable range
    }
}

Usage:

php

<?php
# filename: examples/24-manage-prompt-length.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use App\Services\PromptLengthManager;
use Anthropic\Anthropic;

$manager = new PromptLengthManager();
$longPrompt = str_repeat("This is a very long prompt. ", 10000);

$optimized = $manager->truncateIfNeeded($longPrompt);
$maxTokens = $manager->getRecommendedMaxTokens($optimized);

echo "Original length: " . strlen($longPrompt) . " chars\n";
echo "Optimized length: " . strlen($optimized) . " chars\n";
echo "Recommended max_tokens: {$maxTokens}\n";

Best Practices

Keep prompts focused: Remove unnecessary words and redundancy
Put instructions first: Most important information at the beginning
Use examples efficiently: 3-5 examples usually sufficient
Monitor token usage: Track costs and optimize long-running applications
Reserve response space: Set max_tokens appropriately based on prompt length
Split when necessary: For very long contexts, consider chunking or summarization

Prompt engineering is an iterative process. Rarely does the first prompt produce perfect results. This section covers a systematic approach to improving prompts.

The iterative refinement process follows this cycle:

Write initial prompt
Test with sample inputs
Analyze results for issues
Refine prompt based on findings
Repeat until quality meets requirements

php

<?php
# filename: src/Services/PromptRefiner.php
declare(strict_types=1);

namespace App\Services;

use Anthropic\Anthropic;

class PromptRefiner
{
    public function __construct(
        private readonly Anthropic $client
    ) {}

    public function refine(
        string $initialPrompt,
        array $testCases,
        int $iterations = 3
    ): string {
        $currentPrompt = $initialPrompt;

        for ($i = 0; $i < $iterations; $i++) {
            echo "Iteration " . ($i + 1) . ":\n";

            // Test current prompt
            $results = $this->testPrompt($currentPrompt, $testCases);

            // Analyze results
            $issues = $this->analyzeResults($results);

            if (empty($issues)) {
                echo "✓ Prompt quality acceptable\n";
                break;
            }

            // Refine based on issues
            $currentPrompt = $this->suggestImprovements($currentPrompt, $issues);
            echo "\n";
        }

        return $currentPrompt;
    }

    private function testPrompt(string $prompt, array $testCases): array
    {
        $results = [];

        foreach ($testCases as $testCase) {
            $response = $this->client->messages()->create([
                'model' => 'claude-sonnet-4-20250514',
                'max_tokens' => 512,
                'messages' => [[
                    'role' => 'user',
                    'content' => $prompt . "\n\nInput: " . $testCase['input']
                ]]
            ]);

            $results[] = [
                'input' => $testCase['input'],
                'expected' => $testCase['expected'],
                'actual' => trim($response->content[0]->text),
                'match' => $this->matches($testCase['expected'], trim($response->content[0]->text))
            ];
        }

        return $results;
    }

    private function analyzeResults(array $results): array
    {
        $issues = [];
        $matchCount = 0;

        foreach ($results as $result) {
            if (!$result['match']) {
                $issues[] = [
                    'input' => $result['input'],
                    'expected' => $result['expected'],
                    'actual' => $result['actual'],
                    'type' => 'mismatch'
                ];
            } else {
                $matchCount++;
            }
        }

        echo "Accuracy: {$matchCount}/" . count($results) . "\n";

        return $issues;
    }

    private function suggestImprovements(string $prompt, array $issues): string
    {
        // Ask Claude to suggest improvements
        $issueSummary = $this->formatIssues($issues);

        $refinementPrompt = <<<PROMPT
Analyze this prompt and suggest improvements based on these test failures:

Current prompt:
{$prompt}

Test failures:
{$issueSummary}

Suggest specific improvements to the prompt that would fix these issues.
Return only the improved prompt, no explanation.
PROMPT;

        $response = $this->client->messages()->create([
            'model' => 'claude-sonnet-4-20250514',
            'max_tokens' => 1024,
            'messages' => [['role' => 'user', 'content' => $refinementPrompt]]
        ]);

        return trim($response->content[0]->text);
    }

    private function formatIssues(array $issues): string
    {
        $formatted = [];
        foreach ($issues as $issue) {
            $formatted[] = "Input: {$issue['input']}\nExpected: {$issue['expected']}\nGot: {$issue['actual']}";
        }
        return implode("\n\n", $formatted);
    }

    private function matches(string $expected, string $actual): bool
    {
        // Simple matching - can be enhanced
        return strtolower(trim($expected)) === strtolower(trim($actual));
    }
}

Usage:

php

<?php
# filename: examples/25-iterative-refinement.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use App\Services\PromptRefiner;
use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

$refiner = new PromptRefiner($client);

$initialPrompt = "Classify sentiment: positive or negative";

$testCases = [
    ['input' => 'Great product!', 'expected' => 'positive'],
    ['input' => 'Terrible service', 'expected' => 'negative'],
    ['input' => 'It\'s okay', 'expected' => 'neutral'], // This will fail initially
];

$refinedPrompt = $refiner->refine($initialPrompt, $testCases, iterations: 3);

echo "\nFinal refined prompt:\n{$refinedPrompt}\n";

1. Add Missing Context

php

// Initial (too vague)
$v1 = "Review this code";

// Refined (adds context)
$v2 = "You are a security expert. Review this PHP code for SQL injection vulnerabilities.";

2. Clarify Output Format

php

// Initial (unclear format)
$v1 = "List the issues";

// Refined (specific format)
$v2 = "List security issues as JSON array: [{\"issue\": \"description\", \"severity\": \"high|medium|low\"}]";

3. Add Examples

php

// Initial (no examples)
$v1 = "Classify sentiment";

// Refined (with examples)
$v2 = "Classify sentiment:\n\nExamples:\n\"Great!\" → positive\n\"Terrible\" → negative\n\nClassify: {text}";

4. Fix Edge Cases

php

// Initial (misses edge cases)
$v1 = "Extract email addresses";

// Refined (handles edge cases)
$v2 = "Extract email addresses. Handle formats: user@domain.com, user+tag@domain.co.uk. Return empty array if none found.";

Tracking Prompt Versions

php

<?php
# filename: src/Services/PromptVersionTracker.php
declare(strict_types=1);

namespace App\Services;

class PromptVersionTracker
{
    private array $versions = [];

    public function saveVersion(string $prompt, array $metrics = []): string
    {
        $version = [
            'id' => uniqid('v', true),
            'prompt' => $prompt,
            'timestamp' => time(),
            'metrics' => $metrics,
            'length' => strlen($prompt),
            'estimated_tokens' => (int) ceil(strlen($prompt) / 4)
        ];

        $this->versions[] = $version;
        return $version['id'];
    }

    public function getVersion(string $id): ?array
    {
        foreach ($this->versions as $version) {
            if ($version['id'] === $id) {
                return $version;
            }
        }
        return null;
    }

    public function compareVersions(string $id1, string $id2): array
    {
        $v1 = $this->getVersion($id1);
        $v2 = $this->getVersion($id2);

        if (!$v1 || !$v2) {
            throw new \InvalidArgumentException('Version not found');
        }

        return [
            'length_diff' => $v2['length'] - $v1['length'],
            'token_diff' => $v2['estimated_tokens'] - $v1['estimated_tokens'],
            'metrics_diff' => $this->compareMetrics($v1['metrics'], $v2['metrics'])
        ];
    }

    private function compareMetrics(array $m1, array $m2): array
    {
        $diff = [];
        foreach ($m2 as $key => $value) {
            if (isset($m1[$key])) {
                $diff[$key] = $value - $m1[$key];
            } else {
                $diff[$key] = $value;
            }
        }
        return $diff;
    }

    public function getLatestVersion(): ?array
    {
        return end($this->versions) ?: null;
    }
}

Prompt Testing Methodology

Systematic testing ensures your prompts work reliably across different inputs. This section covers testing strategies and metrics.

Testing Framework

php

<?php
# filename: src/Services/PromptTester.php
declare(strict_types=1);

namespace App\Services;

use Anthropic\Anthropic;

class PromptTester
{
    public function __construct(
        private readonly Anthropic $client
    ) {}

    public function test(
        string $prompt,
        array $testCases,
        array $options = []
    ): array {
        $results = [];
        $temperature = $options['temperature'] ?? 0.0;
        $runs = $options['runs_per_case'] ?? 1;

        foreach ($testCases as $testCase) {
            $caseResults = [];

            for ($i = 0; $i < $runs; $i++) {
                $response = $this->client->messages()->create([
                    'model' => 'claude-sonnet-4-20250514',
                    'max_tokens' => $options['max_tokens'] ?? 512,
                    'temperature' => $temperature,
                    'messages' => [[
                        'role' => 'user',
                        'content' => $this->formatPrompt($prompt, $testCase['input'])
                    ]]
                ]);

                $caseResults[] = [
                    'response' => trim($response->content[0]->text),
                    'tokens' => $response->usage->inputTokens + $response->usage->outputTokens,
                    'latency_ms' => $this->estimateLatency($response)
                ];
            }

            $results[] = [
                'input' => $testCase['input'],
                'expected' => $testCase['expected'] ?? null,
                'runs' => $caseResults,
                'consistency' => $this->checkConsistency($caseResults),
                'accuracy' => $this->checkAccuracy($caseResults, $testCase['expected'] ?? null)
            ];
        }

        return [
            'results' => $results,
            'summary' => $this->generateSummary($results)
        ];
    }

    private function formatPrompt(string $prompt, string $input): string
    {
        return str_replace('{input}', $input, $prompt);
    }

    private function checkConsistency(array $runs): float
    {
        if (count($runs) < 2) {
            return 1.0;
        }

        $responses = array_map(fn($r) => $r['response'], $runs);
        $unique = count(array_unique($responses));

        return 1.0 - (($unique - 1) / count($runs));
    }

    private function checkAccuracy(array $runs, ?string $expected): ?float
    {
        if ($expected === null) {
            return null;
        }

        $matches = 0;
        foreach ($runs as $run) {
            if ($this->matches($expected, $run['response'])) {
                $matches++;
            }
        }

        return $matches / count($runs);
    }

    private function matches(string $expected, string $actual): bool
    {
        return strtolower(trim($expected)) === strtolower(trim($actual));
    }

    private function estimateLatency($response): int
    {
        // Estimate based on response time (would need actual timing in production)
        return 0;
    }

    private function generateSummary(array $results): array
    {
        $totalCases = count($results);
        $consistentCases = 0;
        $accurateCases = 0;
        $totalTokens = 0;

        foreach ($results as $result) {
            if ($result['consistency'] >= 0.8) {
                $consistentCases++;
            }
            if ($result['accuracy'] !== null && $result['accuracy'] >= 0.8) {
                $accurateCases++;
            }
            foreach ($result['runs'] as $run) {
                $totalTokens += $run['tokens'];
            }
        }

        return [
            'total_cases' => $totalCases,
            'consistency_rate' => $consistentCases / $totalCases,
            'accuracy_rate' => $accurateCases / $totalCases,
            'avg_tokens_per_case' => $totalTokens / ($totalCases * count($results[0]['runs']))
        ];
    }
}

Usage:

php

<?php
# filename: examples/26-prompt-testing.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use App\Services\PromptTester;
use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

$tester = new PromptTester($client);

$prompt = "Classify sentiment as positive or negative:\n\nInput: {input}\nOutput:";

$testCases = [
    ['input' => 'Great product!', 'expected' => 'positive'],
    ['input' => 'Terrible service', 'expected' => 'negative'],
    ['input' => 'It works fine', 'expected' => 'positive'], // Ambiguous
];

$results = $tester->test($prompt, $testCases, [
    'runs_per_case' => 3,
    'temperature' => 0.0
]);

echo "Test Summary:\n";
print_r($results['summary']);

echo "\nDetailed Results:\n";
foreach ($results['results'] as $result) {
    echo "Input: {$result['input']}\n";
    echo "Consistency: " . ($result['consistency'] * 100) . "%\n";
    echo "Accuracy: " . ($result['accuracy'] * 100 ?? 'N/A') . "%\n\n";
}

Key Testing Metrics

Accuracy: Percentage of correct outputs
Consistency: How often same input produces same output
Latency: Response time
Token Usage: Cost per request
Edge Case Handling: Performance on unusual inputs

A/B Testing Prompts

php

<?php
# filename: examples/27-ab-testing.php
declare(strict_types=1);

require __DIR__ . '/../vendor/autoload.php';

use App\Services\PromptTester;
use Anthropic\Anthropic;

$client = Anthropic::factory()
    ->withApiKey(getenv('ANTHROPIC_API_KEY'))
    ->make();

$tester = new PromptTester($client);

// Version A: Simple prompt
$promptA = "Classify: {input}";

// Version B: Detailed prompt with examples
$promptB = <<<PROMPT
Classify sentiment as positive or negative.

Examples:
"Great!" → positive
"Terrible" → negative

Classify: {input}
PROMPT;

$testCases = [
    ['input' => 'Amazing product!', 'expected' => 'positive'],
    ['input' => 'Very disappointed', 'expected' => 'negative'],
];

$resultsA = $tester->test($promptA, $testCases);
$resultsB = $tester->test($promptB, $testCases);

echo "Version A Accuracy: " . ($resultsA['summary']['accuracy_rate'] * 100) . "%\n";
echo "Version B Accuracy: " . ($resultsB['summary']['accuracy_rate'] * 100) . "%\n";

if ($resultsB['summary']['accuracy_rate'] > $resultsA['summary']['accuracy_rate']) {
    echo "\n✓ Version B performs better\n";
} else {
    echo "\n✓ Version A performs better\n";
}

Common Pitfalls and Solutions

Pitfall 1: Ambiguous Instructions

php

<?php
# ✗ Bad: Ambiguous
$bad = "Make this code better.";

# ✓ Good: Specific
$good = "Refactor this code to: 1) Add type hints, 2) Extract magic numbers to constants, 3) Add error handling for null values.";

Pitfall 2: Missing Context

php

<?php
# ✗ Bad: No context
$bad = "Is this secure?";

# ✓ Good: With context
$good = "Review this user authentication function for security vulnerabilities. The app handles financial transactions and must comply with PCI DSS.";

Pitfall 3: Unclear Output Format

php

<?php
# ✗ Bad: Vague format
$bad = "List the benefits.";

# ✓ Good: Specific format
$good = "List 5 benefits of using PHP 8.4. Format as JSON array with fields: benefit (string), impact (string), example (string).";

Pitfall 4: Too Many Instructions

php

<?php
# ✗ Bad: Overwhelming
$bad = "Analyze the code, identify bugs, refactor it, write tests, generate documentation, suggest performance improvements, check security, and create a deployment guide.";

# ✓ Good: Focused
$good = "Analyze this code for security vulnerabilities. Focus specifically on: SQL injection, XSS, CSRF, and authentication issues.";

Pitfall 5: Inconsistent Examples

php

<?php
# ✗ Bad: Inconsistent examples
$bad = <<<PROMPT
Examples:
Input: "Great product!" → positive
Input: "It's terrible" → Output: negative
Input: "Okay" -> Classification: neutral

Classify: "I love it"
PROMPT;

# ✓ Good: Consistent format
$good = <<<PROMPT
Examples:
Input: "Great product!"
Output: positive

Input: "It's terrible"
Output: negative

Input: "Okay"
Output: neutral

Classify this:
Input: "I love it"
Output:
PROMPT;

Exercises

Exercise 1: Dynamic Prompt Builder

Build a flexible prompt building system:

php

<?php
class DynamicPromptBuilder
{
    public function forTask(string $task): self
    {
        // TODO: Set task type
        return $this;
    }

    public function withContext(string $context): self
    {
        // TODO: Add context
        return $this;
    }

    public function withExamples(array $examples): self
    {
        // TODO: Add few-shot examples
        return $this;
    }

    public function withFormat(string $format): self
    {
        // TODO: Specify output format
        return $this;
    }

    public function build(): string
    {
        // TODO: Construct final prompt
    }
}

Exercise 2: Prompt Validator

Create a validator to check prompt quality:

php

<?php
class PromptValidator
{
    public function validate(string $prompt): array
    {
        // TODO: Check for clarity
        // TODO: Check for specificity
        // TODO: Check for context
        // TODO: Check for output format specification
        // TODO: Return validation results with suggestions
    }
}

Exercise 3: A/B Prompt Tester

Build a system to test prompt variations:

php

<?php
class PromptABTester
{
    public function test(array $prompts, string $input): array
    {
        // TODO: Run each prompt variation
        // TODO: Collect responses
        // TODO: Compare quality metrics
        // TODO: Return best performing prompt
    }
}

Solution Hints

Exercise 1: Use builder pattern with method chaining, store components in properties, assemble in build() method using string concatenation or template engine.

Exercise 2: Check for question marks, specific requirements, word count, presence of examples. Provide specific improvement suggestions for each issue found.

Exercise 3: Send same input to Claude multiple times with different prompts, measure response length/quality, compare token usage, use Claude to judge which response is better.

Troubleshooting

Inconsistent responses?

Lower temperature (0.0-0.3) for deterministic outputs
Add more specific instructions
Use few-shot examples to guide behavior
Check for ambiguous language in prompt

Response doesn't match format?

Explicitly state format requirements
Provide example of desired output
Add "Return only [format], no explanation"
Use delimiters to separate instructions from data

Claude refuses to follow instructions?

Check if instructions conflict with safety guidelines
Rephrase to be clearer and more specific
Break complex tasks into smaller steps
Provide context for why you need the output

Poor quality outputs?

Add relevant context
Use appropriate model (Opus for complex tasks)
Implement chain-of-thought reasoning
Provide few-shot examples

Wrap-up

Congratulations! You've completed a comprehensive guide to prompt engineering for Claude. Here's what you've accomplished:

✓ Mastered prompt engineering fundamentals including clarity, context, and format specification
✓ Built reusable prompt components including FewShotClassifier, ChainOfThoughtSolver, and PersonaBuilder
✓ Created prompt template systems with PromptTemplate and PromptLibrary for consistent prompting
✓ Learned advanced techniques including few-shot learning (with negative examples), chain-of-thought reasoning, and role-playing
✓ Implemented structured output generation for JSON, XML, CSV, and Markdown formats
✓ Optimized prompt length and learned token management strategies
✓ Mastered iterative refinement with systematic testing and version tracking
✓ Built testing frameworks for evaluating prompt quality and consistency
✓ Identified common pitfalls and learned how to avoid them in your prompts

Key Concepts Learned

Prompt Quality: Clear, specific, and context-rich prompts produce significantly better results
Few-Shot Learning: Providing examples (both positive and negative) guides Claude's responses and improves accuracy
Chain-of-Thought: Encouraging step-by-step reasoning helps with complex problem-solving
Personas: Assigning roles and expertise improves response quality for specialized tasks
Structured Outputs: Explicit format requirements ensure consistent, parseable responses
Templates: Reusable prompt templates save time and ensure consistency across applications
Token Management: Optimizing prompt length reduces costs and improves response quality
Iterative Refinement: Systematic testing and refinement cycles improve prompt effectiveness
Testing Methodology: Proper metrics and A/B testing validate prompt improvements

Next Steps

You now have the foundation to write effective prompts that consistently produce high-quality results. In the next chapter, you'll learn about streaming responses, which allow you to display Claude's output in real-time as it's generated, creating a more interactive user experience.

Key Takeaways

✓ Clarity is crucial - Be specific about what you want
✓ Context matters - Provide relevant background information
✓ Examples guide - Few-shot learning (including negative examples) improves accuracy
✓ Format explicitly - Always specify desired output format
✓ Roles work - Persona assignment improves responses
✓ Chain-of-thought - Helps with complex reasoning tasks
✓ Templates save time - Reusable prompts ensure consistency
✓ Optimize length - Shorter prompts are often more effective and cost less
✓ Iterate systematically - Test, analyze, refine, repeat
✓ Test thoroughly - Use proper metrics and A/B testing to validate improvements
✓ Track versions - Monitor prompt changes and measure impact
✓ Delimiters help - Separate instructions from data clearly
✓ Temperature matters - Lower for consistency, higher for creativity

💻 Code Samples

All code examples from this chapter are available in the GitHub repository:

View Chapter 05 Code Samples

Clone and run locally:

bash

git clone https://github.com/dalehurley/codewithphp.git
cd codewithphp/code/claude-php/chapter-05
composer install
cp .env.example .env
# Add your API key to .env
php examples/01-clarity.php

Chapter 05: Prompt Engineering Basics ​

Overview ​

Prerequisites ​

What You'll Build ​

Objectives ​

Principles of Effective Prompts ​

1. Be Clear and Specific ​

2. Provide Context ​

3. Specify Output Format ​

4. Use Delimiters for Clarity ​

Few-Shot Learning ​

Zero-Shot (No Examples) ​

One-Shot (One Example) ​

Few-Shot (Multiple Examples) ​

Few-Shot with Structured Output ​

Negative Examples (What NOT to Do) ​

Chain-of-Thought Reasoning ​

Basic Chain-of-Thought ​

Explicit Step-by-Step Instructions ​

Self-Consistency Chain-of-Thought ​

Role-Playing and Personas ​

Basic Role Assignment ​

Multi-Attribute Personas ​

Persona Builder Pattern ​

Structured Output Formats ​

JSON Output ​

XML Output ​

CSV Output ​

Markdown Output ​

Prompt Templates ​

Template Engine ​

Prompt Library ​

Prompt Length and Token Management ​

Understanding Token Limits ​

Estimating Prompt Length ​

Optimizing Prompt Length ​

Managing Long Contexts ​

Best Practices ​

Iterative Prompt Refinement ​

The Refinement Cycle ​

Refinement Strategies ​

Tracking Prompt Versions ​

Prompt Testing Methodology ​

Testing Framework ​

Key Testing Metrics ​

A/B Testing Prompts ​

Common Pitfalls and Solutions ​

Pitfall 1: Ambiguous Instructions ​

Pitfall 2: Missing Context ​

Pitfall 3: Unclear Output Format ​

Pitfall 4: Too Many Instructions ​

Pitfall 5: Inconsistent Examples ​

Exercises ​

Exercise 1: Dynamic Prompt Builder ​

Exercise 2: Prompt Validator ​

Exercise 3: A/B Prompt Tester ​

Troubleshooting ​

Wrap-up ​

Key Concepts Learned ​

Next Steps ​

Key Takeaways ​

Further Reading ​

💻 Code Samples ​

Chapter 05: Prompt Engineering Basics

Overview

Prerequisites

What You'll Build

Objectives

Principles of Effective Prompts

1. Be Clear and Specific

2. Provide Context

3. Specify Output Format

4. Use Delimiters for Clarity

Few-Shot Learning

Zero-Shot (No Examples)

One-Shot (One Example)

Few-Shot (Multiple Examples)

Few-Shot with Structured Output

Negative Examples (What NOT to Do)

Chain-of-Thought Reasoning

Basic Chain-of-Thought

Explicit Step-by-Step Instructions

Self-Consistency Chain-of-Thought

Role-Playing and Personas

Basic Role Assignment

Multi-Attribute Personas

Persona Builder Pattern

Structured Output Formats

JSON Output

XML Output

CSV Output

Markdown Output

Prompt Templates

Template Engine

Prompt Library

Prompt Length and Token Management

Understanding Token Limits

Estimating Prompt Length

Optimizing Prompt Length

Managing Long Contexts

Best Practices

Iterative Prompt Refinement

The Refinement Cycle

Refinement Strategies

Tracking Prompt Versions

Prompt Testing Methodology

Testing Framework

Key Testing Metrics

A/B Testing Prompts

Common Pitfalls and Solutions

Pitfall 1: Ambiguous Instructions

Pitfall 2: Missing Context

Pitfall 3: Unclear Output Format

Pitfall 4: Too Many Instructions

Pitfall 5: Inconsistent Examples

Exercises

Exercise 1: Dynamic Prompt Builder

Exercise 2: Prompt Validator

Exercise 3: A/B Prompt Tester

Troubleshooting

Wrap-up

Key Concepts Learned

Next Steps

Key Takeaways

Further Reading

💻 Code Samples