Overview¶

LLMling's prompt library allows defining reusable system prompts that can be shared across agents. Prompts are defined in the prompts section of your configuration and can be referenced by name.

System prompts define agent behavior, personality, and methodology. You can configure prompts using:

Static prompts: Inline text content
File prompts: Load from external files with Jinja2 templating
Library prompts: Reference shared prompts from the library
Function prompts: Dynamically generate prompts using Python functions

Basic Structure¶

# yaml-language-server: $schema=https://raw.githubusercontent.com/phil65/llmling-agent/refs/heads/main/schema/config-schema.json
prompts:
  # Define reusable system prompts
  system_prompts:
    expert_analyst:
      content: |
        You are an expert data analyst.
        Focus on finding patterns and insights.
      category: role

    step_by_step:
      content: |
        Break tasks into sequential steps.
        Explain each step thoroughly.
      category: methodology

# Using prompts in agents
agents:
  analyst:
    system_prompts:
      # Direct string prompts
      - "You help with analysis."

      # Reference library prompts
      - type: library
        reference: expert_analyst
      - type: library
        reference: step_by_step

Prompt Categories¶

System prompts can be categorized by their purpose:

Role: Define WHO the agent is (e.g., "expert developer", "data scientist")
Methodology: Define HOW the agent works (e.g., "step-by-step", "analytical")
Tone: Define communication STYLE (e.g., "professional", "friendly")
Format: Define output STRUCTURE (e.g., "markdown", "structured")

Configuration Reference¶

StaticPromptConfig¶

Configuration for a static text prompt.

StaticPromptConfig (YAML)
- type: static
  content: You are a helpful assistant  # The prompt text content.
  category: role  # Categorization for template organization.

FilePromptConfig¶

Configuration for a file-based Jinja template prompt.

FilePromptConfig (YAML)
- type: file
  path: prompts/system.j2  # Path to the Jinja template file.
  variables: {}  # Variables to pass to the template.
  content: You are a helpful assistant  # The actual prompt text.
  category: role  # Categorization for template organization.

LibraryPromptConfig¶

Configuration for a library reference prompt.

LibraryPromptConfig (YAML)
- type: library
  reference: coding_assistant  # Library prompt reference identifier.

FunctionPromptConfig¶

Configuration for a function-generated prompt.

FunctionPromptConfig (YAML)
- type: function
  function: mymodule.generate_prompt  # Import path to the function that generates the prompt.
  arguments: {}  # Arguments to pass to the function.
  content: You are a helpful assistant  # The actual prompt text.
  category: role  # Categorization for template organization.

Complete Example¶

prompts:
  system_prompts:
    # Role definitions
    technical_expert:
      category: role
      content: |
        You are a technical expert specializing in:
        - Software development best practices
        - System architecture and design
        - Code review and quality assurance

    # Methodology definitions
    systematic:
      category: methodology
      content: |
        Follow this systematic approach:
        1. Understand requirements fully
        2. Break down complex problems
        3. Apply best practices consistently
        4. Validate results thoroughly

    # Tone definitions
    professional:
      category: tone
      content: |
        Maintain professional communication:
        - Use formal, precise language
        - Be respectful and constructive
        - Provide clear explanations

agents:
  senior_dev:
    model: gpt-4
    system_prompts:
      - "Specialize in Python and TypeScript development."
      - type: library
        reference: technical_expert
      - type: library
        reference: systematic
      - type: library
        reference: professional
      - type: file
        path: "prompts/coding_style.j2"

Organization Best Practices¶

File Structure¶

Keep prompts organized in separate files:

# prompts/roles.yml
prompts:
  system_prompts:
    technical_expert:
      category: role
      content: ...

# prompts/styles.yml
prompts:
  system_prompts:
    professional:
      category: tone
      content: ...

# agents.yml
INHERIT:
  - prompts/roles.yml
  - prompts/styles.yml

agents:
  my_agent:
    system_prompts:
      - type: library
        reference: technical_expert
      - type: library
        reference: professional