prompt-schema

Convert Zod schemas to AI prompts in Markdown.

---

Table of Contents

• What is this?
• Quick Start
• Themes
• Options
• Advanced Usage
• API Reference
• Supported Features
• License

---

What is this?

When using LLMs with structured output (OpenAI function calling, AI SDK generateObject, etc.), you need to tell the model what schema to follow. This library converts your Zod schemas into clear, Markdown-formatted prompt instructions that LLMs understand well.

Before (raw JSON Schema dumped into prompt):

{
  "type": "object",
  "properties": {
    "name": { "type": "string", "maxLength": 100 },
    "email": { "type": "string", "format": "email" },
    "age": { "type": "number", "minimum": 0, "maximum": 120 }
  },
  "required": ["name", "email", "age"]
}

After (prompt-schema output):

## Schema

- name: string (required • max 100 chars)
- email: string (required • format: email)
- age: number (required • min: 0 • max: 120)

---

Quick Start

Install

npm install prompt-schema zod

Requires zod@^3.25.0 (includes both v3 and v4 APIs)

Basic Usage

import { getPrompts } from 'prompt-schema';
import { z } from 'zod/v3';

const schema = z.object({
  name: z.string().max(100),
  email: z.string().email(),
  age: z.number().min(0).max(120),
});

const prompt = getPrompts(schema);

Output:

## Schema

- name: string (required • max 100 chars)
- email: string (required • format: email)
- age: number (required • min: 0 • max: 120)

---

Themes

Choose an output format that fits your use case. The standard, expanded, and condensed themes output Markdown, which LLMs parse reliably:

Standard (default)

Clean and readable, optimized for AI prompts:

## Schema

- name: string (required)
- age: number (required • min: 0)
- email: string (required • format: email)

Condensed

Ultra-compact for token efficiency:

name:str!
age:num![≥0]
email:str![@]

Expanded

Includes full constraint details:

## Schema

- name: string (required)
- age: number (required • min: 0)
- email: string (required • pattern: ^[...] • format: email)

## Examples

{ "name": "John Doe", "age": 30, "email": "john@example.com" }

JSON

Structured data for programmatic use:

{
  "schema": {
    "fields": [
      { "name": "name", "type": "string", "required": true },
      { "name": "age", "type": "number", "required": true, "constraints": { "min": 0 } },
      { "name": "email", "type": "string", "required": true, "constraints": { "format": "email" } }
    ]
  }
}

Using Themes

const prompt = getPrompts(schema, { theme: 'condensed' });

---

Options

Option	Type	Default	Description
theme	'standard' | 'condensed' | 'expanded' | 'json'	'standard'	Output format
maxDepth	number	3	Maximum nesting depth for objects/arrays
safe	boolean	false	Return fallback message on error instead of throwing

---

Advanced Usage

Zod v4 with Examples

Zod v4 supports .meta() for adding examples:

import { getPrompts } from 'prompt-schema';
import { z } from 'zod/v4';

const schema = z
  .object({
    username: z.string().min(3).max(20),
    age: z.number().min(18),
  })
  .meta({
    examples: [{ username: 'john_doe', age: 25 }],
  });

const prompt = getPrompts(schema, { theme: 'expanded' });

Complex Types

import { z } from 'zod/v3';

// Unions
const unionSchema = z.object({
  value: z.union([z.string(), z.number(), z.boolean()]),
});
// → value: string | number | boolean (required)

// Tuples
const tupleSchema = z.object({
  coordinates: z.tuple([z.number(), z.number()]),
});
// → coordinates: [number, number] (required)

// Records
const recordSchema = z.object({
  metadata: z.record(z.string(), z.string()),
});
// → metadata: Record<string, string> (required)

// Discriminated Unions
const notificationSchema = z.object({
  notification: z.discriminatedUnion('type', [
    z.object({ type: z.literal('email'), to: z.string().email() }),
    z.object({ type: z.literal('sms'), phone: z.string() }),
  ]),
});

Custom Adapters

For advanced use cases, use the PromptSchema class directly:

import { PromptSchema, zodV3Adapter, zodV4Adapter, jsonSchemaAdapter } from 'prompt-schema';

const converter = new PromptSchema();
converter.registerAdapter(zodV4Adapter);
converter.registerAdapter(zodV3Adapter);
converter.registerAdapter(jsonSchemaAdapter);

const prompt = converter.toPrompt(schema, { theme: 'expanded' });

See CONTRIBUTING.md for creating custom adapters.

---

API Reference

getPrompts(schema, options?)

Main entry point. Auto-detects Zod v3 or v4.

function getPrompts(schema: unknown, options?: ContextOptions): string;

PromptSchema Class

class PromptSchema {
  registerAdapter(adapter: SchemaAdapter): void;
  toPrompt(schema: unknown, options?: ContextOptions): string;
  getAdapters(): string[];
}

Types

interface ContextOptions {
  theme?: 'standard' | 'condensed' | 'expanded' | 'json';
  maxDepth?: number;
  safe?: boolean;
}

interface SchemaAdapter<T = unknown> {
  name: string;
  canHandle: (schema: unknown) => boolean;
  toJsonSchema: (schema: T, options?: unknown) => JsonSchema;
}

---

Supported Features

Fully Supported

• Objects with nested properties
• Arrays with typed items
• Tuples (fixed-length arrays)
• Records (dynamic key-value maps)
• Enums and literal values
• Optional and nullable fields
• Discriminated unions
• Regular unions
• Date types (z.date() → format: date-time)
• All Zod constraints (min, max, length, regex, format)
• Zod v3 and v4

Partially Supported

• Intersections - Treated as merged objects, may lose some type information

Not Supported

These Zod features don't translate to JSON Schema:

• transform(), refine() - Runtime-only validations
• lazy() - Recursive schemas would cause infinite loops
• set(), map() - No JSON Schema equivalent
• bigint(), symbol(), function(), promise() - Not serializable
• never(), unknown(), any(), void(), undefined(), nan() - Fall back to generic types

---

License

MIT - Created by Joe Seifi