Prompt Development Guide

This guide provides comprehensive information about the prompt template system used in Nexent for creating different types of agents. The YAML files in the backend/prompts/ directory define system prompts, planning prompts, and other key prompt components for various agent types.

File Naming Convention

The naming format follows {agent_type}_agent.yaml, where:

• agent_type: Describes the main function or purpose of the agent (e.g., manager, search, etc.)

Prompt Template Structure

Each YAML file contains the following main sections:

1. system_prompt

The system prompt is the core component of an agent, defining its role, capabilities, and behavioral guidelines. It typically includes:

• Core Responsibilities: Main duties and capabilities description
• Execution Flow: Standard process and methods for agent task execution
• Available Resources: List of tools and sub-agents the agent can use
• Resource Usage Requirements: Priority and strategy for using different tools
• Python Code Standards: Code writing standards and constraints
• Example Templates: Examples demonstrating agent task execution

2. planning

Contains various prompts for task planning:

• initial_facts: Initial fact collection prompts
• initial_plan: Initial plan formulation prompts
• update_facts_pre_messages: Prompts before updating facts
• update_facts_post_messages: Prompts after updating facts
• update_plan_pre_messages: Prompts before updating plans
• update_plan_post_messages: Prompts after updating plans

3. managed_agent

Defines prompts for sub-agent interactions:

• task: Task assignment prompts for sub-agents
• report: Prompts for sub-agent result reporting

4. final_answer

Defines prompts for final answer generation:

• pre_messages: Prompts before generating final answers
• post_messages: Prompts after generating final answers

5. tools_requirement

Defines prompts for tool usage standards and priorities.

6. few_shots

Provides few-shot learning examples to help agents better understand task execution methods.

Template Variables

The prompt templates use the following special variables for dynamic replacement:

• : Available tools list
• : Available sub-agents list
• : Current task description
• : Authorized Python module imports
• : Updated facts list
• : Known facts list
• : Remaining execution steps

Available Prompt Templates

Core Templates

1. Manager Agent Templates

   • manager_system_prompt_template.yaml - Chinese version
   • manager_system_prompt_template_en.yaml - English version

   These templates define the core manager agent that coordinates and dispatches various assistants and tools to efficiently solve complex tasks.

2. Managed Agent Templates

   • managed_system_prompt_template.yaml - Chinese version
   • managed_system_prompt_template_en.yaml - English version

   These templates define specialized agents that perform specific tasks under the coordination of the manager agent.

3. Specialized Agent Templates

   • knowledge_summary_agent.yaml - Knowledge summary agent (Chinese)
   • knowledge_summary_agent_en.yaml - Knowledge summary agent (English)
   • analyze_file.yaml - File analysis agent (Chinese)
   • analyze_file_en.yaml - File analysis agent (English)

Utility Templates

Located in the utils/ directory:

1. Prompt Generation Templates

   • prompt_generate.yaml - Chinese version
   • prompt_generate_en.yaml - English version

   These templates help generate efficient and clear prompts for different agent types.

2. Prompt Fine-tuning Templates

   • prompt_fine_tune.yaml - Chinese version
   • prompt_fine_tune_en.yaml - English version

   Templates for fine-tuning and optimizing existing prompts.

3. Title Generation Templates

   • generate_title.yaml - For generating titles and summaries

Execution Flow

The standard agent execution flow follows this pattern:

1. Think: Analyze current task status and progress
2. Code: Write simple Python code following standards
3. Observe: View code execution results
4. Repeat: Continue the cycle until the task is complete

Code Standards

When writing Python code in prompts:

1. Use the format 代码：\n```py\n for executable code
2. Use the format 代码：\n```code:language_type\n for display-only code
3. Use only defined variables that persist across calls
4. Use print() function to make variable information visible
5. Use keyword parameters for tool and agent calls
6. Avoid excessive tool calls in a single round
7. Only import from authorized modules:

Best Practices

1. Task Decomposition: Break complex tasks into manageable sub-tasks
2. Professional Matching: Assign tasks based on agent expertise
3. Information Integration: Integrate outputs from different agents
4. Efficiency Optimization: Avoid redundant work
5. Result Evaluation: Assess agent return results and provide additional guidance when needed

Example Usage

Here's an example of how a manager agent might coordinate with specialized agents:

# Example task assignment
managed_agent:
  task: |
    Please analyze the provided document and extract key insights.
    
  report: |
    {{final_answer}}

This system allows for flexible and powerful agent coordination while maintaining clear standards and best practices for prompt development.