Write Code From Spec

This example shows how to create a simple LLM function that generates TypeScript code from a plain-text specification. It's ideal when you want a quick implementation based on a natural language prompt.

---

Step 1 – Prepare the Prompt

We define the LLM instructions using a templated string. The prompt tells the LLM to act as an expert developer and return only TypeScript code, no prose.

const PROMPT = `You are a senior TypeScript developer. 

Write a concise implementation for the task: "{{spec}}". 

Respond with only the code (no questions or explanation), inside a single ts code block`;

Tip: You want to instruct the llm to respond with the markdown code block - this improves the reliability of code extraction with the parser.

---

Step 2 – Define the Function

This function configures the LLM, applies the prompt, and parses the response using markdownCodeBlock. It wraps everything into a single executor.

export async function writeCodeFromSpec(spec: string) {
  // define the LLM
  const llm = useLlm("openai.gpt-4o-mini"); // change this if you want
  // instantiate the prompt
  const prompt = createChatPrompt<{ spec: string }>(PROMPT);
  // define the output parser
  const parser = createParser("markdownCodeBlock");
  // tie everything together in an executor
  const executor = createLlmExecutor({ llm, prompt, parser });

  return executor.execute({ spec });
}

Why this setup? By separating the prompt and parser setup, the function remains readable and testable. It also allows you to swap in new prompts or models easily.

---

Step 3 – Use it

Here’s how you can call the writeCodeFromSpec function to generate code dynamically:

import { writeCodeFromSpec } from "./your-path";

const code = await writeCodeFromSpec("a function that adds two numbers");

console.log(code);
/**
 * ```ts
 * { "code": "function add(a: number, b: number): number {\nreturn a + b;\n", "language": "typescript" }
 * ```
 */

---

Complete File

import {
  createChatPrompt,
  createParser,
  createLlmExecutor,
  useLlm,
} from "llm-exe";
const PROMPT = `You are a senior TypeScript developer. 

Write a concise implementation for the task: "{{spec}}". 

Respond with only the code (no questions or explanation), inside a single ts code block`;

export async function writeCodeFromSpec(spec: string) {
  // define the LLM
  const llm = useLlm("openai.gpt-4o-mini"); // change this if you want
  // instantiate the prompt
  const prompt = createChatPrompt<{ spec: string }>(PROMPT);
  // define the output parser
  const parser = createParser("markdownCodeBlock");
  // tie everything together in an executor
  const executor = createLlmExecutor({ llm, prompt, parser });

  return executor.execute({ spec });
}