Skip to content

PromptBuilder Guide

The PromptBuilder provides a structured way to construct LLM prompts with consistent formatting and content organization. It handles content rendering, instruction management, and context tracking.

Key Features

  • Content Type System: Predefined content types for common prompt elements
  • Custom Renderers: Add your own content renderers for specific needs
  • Instruction Management: Easily add and organize instructions
  • Context Tracking: Maintain conversation and execution context
  • Step Result Handling: Specialized rendering for different step types

Basic Usage

  1. Create a new PromptBuilder instance: 

    const promptBuilder = new PromptBuilder();

  2. Add content using predefined content types: 

    promptBuilder.addContent(ContentType.ARTIFACTS, artifacts);
promptBuilder.addContent(ContentType.CONVERSATION, chatPosts);

  3. Add instructions and context: 

    promptBuilder.addInstruction("You are a helpful assistant");
promptBuilder.addContext("Current date: 2025-01-23");

  4. Build the final prompt: 

    const prompt = promptBuilder.build();

Content Types

The system provides these standard content types:

Type Description
ARTIFACTS Attached files or data artifacts
CONVERSATION Chat history and context
SEARCH_RESULTS Web search or database query results
CODE Source code snippets
DOCUMENTS Long-form text documents
TASKS Task lists and project details
GOALS Project goals and objectives
STEP_RESULTS Previous step execution results
EXECUTE_PARAMS Current execution parameters

Custom Renderers

You can add custom content renderers for specific needs:

promptBuilder.registerRenderer(ContentType.CUSTOM_TYPE, (content) => {
  return `Custom Content:\n${JSON.stringify(content)}`;
});

Step Result Handling

The PromptBuilder includes special handling for different step types:

// Register a custom step result renderer
promptBuilder.registerStepResultRenderer(
  StepResultType.Validation,
  (step) => `Validation Result: ${step.response?.message}`
);

// Add step results
promptBuilder.addContent(ContentType.STEP_RESULTS, steps);

Best Practices

  1. Organize Content: Use appropriate content types to keep prompts structured
  2. Reuse Renderers: Create reusable renderers for common content formats
  3. Maintain Context: Add relevant execution context to each prompt
  4. Use Instructions: Clearly define the LLM's role and constraints
  5. Validate Outputs: Test rendered prompts to ensure proper formatting

Example Usage

Here's a complete example showing how to build a validation prompt:

const promptBuilder = new PromptBuilder();

// Add core instructions
promptBuilder.addInstruction("You are validating whether a proposed solution addresses the original goal");

// Add execution parameters
promptBuilder.addContent(ContentType.EXECUTE_PARAMS, params);

// Add previous results
promptBuilder.addContent(ContentType.STEP_RESULTS, steps);

// Add evaluation guidelines
promptBuilder.addInstruction(`Evaluation Guidelines:
1. Consider if the solution makes reasonable progress
2. Allow for iterative improvement
3. Focus on critical issues`);

// Build the final prompt
const prompt = promptBuilder.build();

This will generate a well-structured prompt with clear sections and consistent formatting.