Introduction

prai (programmable ai) is a JavaScript framework for building step-by-step LLM instructions. It transforms messy natural language prompts into structured, maintainable, and debuggable code, leading to high-quality, reliable outputs.

npm install prai
# or
pnpm add prai

Here's a simple example that generates brand adjectives and creates a color theme:

import { History, Model, openai, step } from 'prai'
import { z } from 'zod'

// Create a model
const model = new Model({
  name: 'gpt-4.1-mini',
  provider: openai({ apiKey: process.env.OPENAI_API_KEY })
})

// Create history to track the workflow
const history = new History()

// Step 1: Generate brand adjectives
const adjectives = await step(
  `List some adjectives fitting the design of the pmndrs brand which is a open source developer collective`,
  z.array(z.string()),
  { model, history }
)

// Step 2: Create a color theme based on the adjectives
const theme = await step(
  `Based on the ${history.reference(adjectives)}, derive a fitting color theme`,
  z.object({
    primary: z.string().describe('hex color'),
    secondary: z.string().describe('hex color'),
    accent: z.string().describe('hex color')
  }),
  { model, history }
)

console.log(theme)
// { primary: "#1a1a1a", secondary: "#f0f0f0", accent: "#ff6b6b" }

Writing prompts in natural language is great—until they become messy with multiple steps, output format descriptions, and too many lines of text. prai solves this by making prompts into code that is:

Steps

Steps are the fundamental building blocks of prai. Each step represents a single instruction to an AI model with a defined input prompt and expected output schema using Zod. This ensures type safety and forces the AI to return data in the exact format you need.

History

History tracks the conversation flow between steps, allowing later steps to reference results from previous steps. This creates a coherent workflow where each step builds upon previous results.

Model

Models represent AI providers (OpenAI, Groq, Gemini) and handle the actual communication with AI services. They abstract away provider-specific details.

We recommend prai for problems that are best solved with a structured process, where some steps involve reasoning or decision-making. Examples include: