Getting Started

Why Structured Prompts?

Every LLM prompt today is a text string. While this works for simple cases, it creates problems as applications grow:

• Brittleness: Changing one part breaks unrelated functionality
• Maintainability: No clear structure or semantic meaning
• Testability: Hard to verify prompt components work as intended
• Composability: Difficult to reuse prompt patterns across projects

Bolt Foundry solves these problems by transforming prompts from strings into composable cards with specifications and examples.

Core Concepts

Cards

Cards are the building blocks of structured prompts. Each card encapsulates:

• Specifications: What the AI should do
• Examples: Scored samples showing good (-3 to +3) behavior
• Context: Required inputs and their descriptions

Specifications

Specifications define the behavior you want from the AI. They're organized hierarchically:

card
  .spec("Main behavior")
  .specs("category", (s) =>
    s.spec("Specific behavior 1")
      .spec("Specific behavior 2"));

Scoring System

The -3 to +3 scoring isn't just about examples - it's about achieving 99% reliability through inference-time control:

• +3: Excellent example of desired behavior
• +2: Good example
• +1: Acceptable
• -1: Mildly undesirable
• -2: Bad example
• -3: Completely wrong behavior

Your First Card

Let's build a customer support assistant step by step:

import { BfClient } from "@bolt-foundry/bolt-foundry";

const client = BfClient.create();

const supportAgent = client.createAssistantCard(
  "customer-support",
  (card) =>
    card
      .spec("You are a helpful customer support agent")
      .specs("personality", (p) =>
        p.spec("Be empathetic and professional", {
          samples: (s) =>
            s.sample("I understand your frustration. Let me help.", 3)
              .sample("Whatever, that's not my problem.", -3),
        }))
      .context((ctx) =>
        ctx.string("customerQuery", "What the customer is asking")
      ),
);

Next Steps

• Check out the Quickstart for a practical example
• Learn about advanced patterns in our guides
• Explore the API reference for detailed documentation