Prompts

Understanding prompts and prompt engineering in PromptKit.

What is a Prompt?

A prompt is the text input sent to an LLM. It tells the model what to do and provides context for generating a response.

Basic Example

User: What's the capital of France?

LLM Response: The capital of France is Paris.

Structured Prompt

System: You are a helpful assistant. Be concise.

User: What's the capital of France?

LLM Response: Paris.

Prompt Components

Messages

Prompts consist of messages with different roles:

System: Instructions for the AI’s behavior
User: Input from the end user
Assistant: Previous AI responses (for context)
Tool: Results from function calls

Example in PromptKit:

messages := []types.Message{
    {Role: "system", Content: "You are a helpful assistant"},
    {Role: "user", Content: "Hello"},
    {Role: "assistant", Content: "Hi! How can I help?"},
    {Role: "user", Content: "What's 2+2?"},
}

System Prompts

System prompts set the AI’s behavior:

Good system prompt:

You are a customer support agent for TechCorp.

Guidelines:
- Be professional and empathetic
- Keep responses under 100 words
- Reference documentation when possible
- Escalate complex issues to human agents

Available resources:
- Help docs: https://help.techcorp.com
- Status page: https://status.techcorp.com

Poor system prompt:

Be helpful.

Prompt Engineering Techniques

1. Clear Instructions

Bad:

Write something about dogs.

Good:

Write a 3-sentence summary about Golden Retrievers.
Include: origin, temperament, and typical uses.

2. Few-Shot Examples

Provide examples to guide the model:

Convert user questions to support ticket categories.

Examples:
Q: "I can't log in" → Category: Authentication
Q: "My credit card was declined" → Category: Billing
Q: "The app crashed" → Category: Technical

Q: "I forgot my password" → Category: ?

3. Chain-of-Thought

Ask the model to explain its reasoning:

Solve this math problem step-by-step:
If Alice has 3 apples and Bob gives her 5 more, how many does Alice have?

Let's solve this step by step:
1. Alice starts with: 3 apples
2. Bob gives her: 5 apples
3. Total: 3 + 5 = 8 apples

Answer: 8 apples

4. Role-Based

Assign a specific role:

You are an experienced software architect reviewing code.
Focus on:
- Design patterns
- Scalability
- Maintainability

Review this code:
[code here]

Prompts in PromptKit

Runtime

Runtime uses prompts with templates:

templates := template.NewRegistry()
templates.RegisterTemplate("support", &template.PromptTemplate{
    SystemPrompt: "You are a support agent. Be helpful.",
    Variables: map[string]string{
        "company": "TechCorp",
    },
})

pipe := pipeline.NewPipeline(
    middleware.TemplateMiddleware(templates, nil),
    middleware.ProviderMiddleware(provider, nil, nil, nil),
)

result, _ := pipe.Execute(ctx, "user", "Help me")

SDK

SDK simplifies prompt management:

conv := sdk.NewConversation(provider, &sdk.ConversationConfig{
    SystemPrompt: "You are a helpful assistant",
    Model:        "gpt-4o-mini",
})

response, _ := conv.Send(ctx, "Hello")

PackC

PackC packages prompts for distribution:

# support.prompt
system: |
  You are a  for .
  Be professional and concise.

user: |

Pack and use:

packc pack prompts/ -o support.pack

pack, _ := packc.LoadPack("support.pack")
template := pack.Templates["support"]

PromptArena

PromptArena tests prompts:

tests:
  - name: Support Query
    prompt: "How do I reset my password?"
    assertions:
      - type: contains
        value: "password reset"
      - type: max_length
        value: 200

Best Practices

Do’s

✅ Be specific - Clear instructions get better results
✅ Provide context - Give the model relevant information
✅ Use examples - Show what you want
✅ Set constraints - Specify format, length, tone
✅ Test variations - Try different phrasings
✅ Version prompts - Track changes over time

Don’ts

❌ Don’t be vague - “Be helpful” isn’t enough
❌ Don’t assume knowledge - Provide necessary context
❌ Don’t skip system prompts - They guide behavior
❌ Don’t ignore length - Longer ≠ better
❌ Don’t forget to test - Use PromptArena

Common Patterns

Task-Oriented

Task: Extract customer email from support ticket

Input: "Hi, my email is john@example.com and I need help"
Output: john@example.com

Input: "Please contact me at support@company.com"
Output:

Conversational

You are a friendly chatbot. Maintain context across messages.

Be:
- Conversational and warm
- Helpful and informative
- Concise (under 50 words per response)

Analytical

Analyze the following text and provide:
1. Main topic
2. Sentiment (positive/negative/neutral)
3. Key entities mentioned
4. Confidence score (0-1)

Text: [input]

Creative

Write a creative product description for:

Product:
Features:
Target audience:

Style: Engaging and persuasive
Length: 50-100 words

Prompt Management

Development

Store prompts in version control
Use meaningful names
Document prompt purpose
Track performance metrics

Testing

Test with PromptArena
Try edge cases
Measure quality
Compare variations

Production

Load from PackC files
Use templates for flexibility
Monitor performance
Version prompts

Troubleshooting

Problem: Inconsistent Responses

Solution: Add more specific instructions and examples

Problem: Wrong Format

Solution: Specify exact format in prompt

Before:

List the items.

After:

List the items in JSON format:
{"items": ["item1", "item2"]}

Problem: Ignoring Instructions

Solution: Emphasize important instructions

IMPORTANT: Always include sources.
NEVER make up information.

[rest of prompt]

Problem: Too Verbose

Solution: Set length constraints

Answer in exactly 2 sentences. Be concise.

Summary

Prompts are the foundation of LLM applications. Good prompts:

Provide clear instructions
Include relevant context
Use examples when helpful
Set appropriate constraints
Are tested and versioned

Templates - Organizing prompts
Validation - Content safety
Runtime Templates - Implementation
PackC - Prompt packaging