Skip to main content

​
Prompting Best Practices

BAML makes it easy to write and manage prompts for structured LLM outputs. This guide covers best practices for crafting effective prompts.

​
Chain-of-Thought Prompting

Chain-of-thought prompting encourages the language model to think step by step, reasoning through the problem before providing an answer. This can improve the quality of responses and make them easier to understand.

​
Technique 1: Reasoning Before Output

Require the model to reason before outputting the structured object: 
function GetOrderInfo(email: Email) -> OrderInfo {
  client "openai/gpt-5-mini"
  prompt #"
    Extract everything from this email.

    {{ ctx.output_format }}

    Before you answer, please explain your reasoning step-by-step. 
    
    For example:
    If we think step by step we can see that ...

    Therefore the output is:
    {
      ... // schema
    }

    {{ _.role('user') }}

    Sender: {{email.from_address}}
    Email Subject: {{email.subject}}
    Email Body: {{email.body}}
  "#
}
Allow the model to outline relevant information without constraining the format: 
function GetOrderInfo(email: Email) -> OrderInfo {
  client "openai/gpt-5-mini"
  prompt #"
    Extract everything from this email.

    {{ ctx.output_format }}

    Outline some relevant information before you answer.
    Example:
    - ...
    - ...
    ...
    {
      ... // schema
    }

    {{ _.role('user') }}

    Sender: {{email.from_address}}
    Email Subject: {{email.subject}}
    Email Body: {{email.body}}
  "#
}
This technique uses - ... to indicate the model should output information without limiting it to a specific format or number of items.

​
Technique 3: Embed Reasoning in Schema

Add reasoning fields directly to your output schema: 
class OrderInfo {
    clues string[] @description(#"
      relevant quotes from the email related to shipping
    "#)
    order_status "ORDERED" | "SHIPPED" | "DELIVERED" | "CANCELLED"
    tracking_number string?
    estimated_arrival_date string?
}

​
Technique 4: Reasoning as Comments

Ask the model to embed reasoning as comments in the structured output: 
class OrderInfo {
    order_status "ORDERED" | "SHIPPED" | "DELIVERED" | "CANCELLED"
      @description(#"
        before fields, in comments list out any relevant clues from the email
      "#)
    tracking_number string?
    estimated_arrival_date string?
}

​
Reusable Prompt Snippets

Use template_string to create reusable prompt components: 
template_string ChainOfThought() #"
    Outline some relevant information before you answer.
    Example:
    - ...
    - ...
    ...
    {
      ... // schema
    }
"#

function GetOrderInfo(email: Email) -> OrderInfo {
  client "openai/gpt-5-mini"
  prompt #"
    Extract everything from this email.

    {{ ctx.output_format }}

    {{ ChainOfThought() }}

    {{ _.role('user') }}

    Sender: {{email.from_address}}
    Email Subject: {{email.subject}}
    Email Body: {{email.body}}
  "#
}

​
Reducing Hallucinations

​
Set Temperature to 0.0

For data extraction tasks, use zero temperature to make the model less creative: 
client<llm> MyClient {
  provider openai
  options {
    model "gpt-5-mini"
    temperature 0.0
  }
}

​
Reduce Input Tokens

Prune unnecessary data to reduce confusion:
  • Remove irrelevant fields from input data
  • Split large prompts into smaller, focused prompts
  • For images, crop unnecessary parts and verify clarity at model resolution

​
Use Clear, Unambiguous Instructions

Avoid contradictions and be explicit about format requirements: 
function ExtractData(input: string) -> DataSchema {
  client "openai/gpt-5-mini"
  prompt #"
    Extract information from the following text.
    
    {{ ctx.output_format }}
    
    It's ok if this isn't fully valid JSON, 
    we will fix it afterwards and remove any comments.
    
    {{ _.role('user') }}
    {{ input }}
  "#
}

​
Token Optimization

Optimize token usage with alternative data formats: 
function AnalyzeProducts(products: Product[]) -> Analysis {
  client GPT4
  prompt #"
    Analyze these products and provide insights:
    
    {{ products|format(type="toon") }}
    
    Focus on pricing trends and inventory status.
  "#
}
BAML supports multiple formats:
  • json - Standard JSON (default)
  • yaml - YAML format
  • toon - Token-Oriented Object Notation (compact)
Always test alternative formats with your specific use case. Lower token count doesn’t guarantee better accuracy or latency.

​
Classification Tasks

For classification tasks, use enums for type safety: 
enum MessageType {
  SPAM
  NOT_SPAM
}

function ClassifyText(input: string) -> MessageType {
  client "openai/gpt-5-mini"
  prompt #"
    Classify the message. 

    {{ ctx.output_format }}

    {{ _.role("user") }} 
    
    {{ input }}
  "#
}
For multi-label classification, use arrays: 
enum TicketLabel {
  ACCOUNT
  BILLING
  GENERAL_QUERY
}

class TicketClassification {
  labels TicketLabel[]
}

function ClassifyTicket(ticket: string) -> TicketClassification {
  client "openai/gpt-5-mini"
  prompt #"
    You are a support agent. Analyze the support ticket and select all applicable labels.

    {{ ctx.output_format }}

    {{ _.role("user") }}
    
    {{ ticket }}
  "#
}

​
Testing Your Prompts

Use BAML’s built-in testing to validate your prompts: 
test BasicTest {
  functions [ClassifyText]
  args {
    input "Buy cheap watches now! Limited time offer!!!"
  }
}

test NonSpamTest {
  functions [ClassifyText]
  args {
    input "Hey Sarah, can we meet at 3 PM tomorrow to discuss the project?"
  }
}
Run tests in the BAML Playground or using the CLI to validate your prompts against real examples.

​
Next Steps

Build docs developers (and LLMs) love

Get started for freeTalk to us