LLM Prompt Engineering: Advanced Techniques for Production AI Applications

### System Prompt Template:
You are an expert TypeScript Code Auditor.
Your task is to analyze user-submitted TypeScript snippets for potential null-pointer bugs and type leaks.

CONSTRAINTS:
- Do NOT generate explanatory conversation or greetings.
- Output MUST be strictly valid JSON matching the following schema.
- If no bugs are found, return an empty "issues" array.

OUTPUT JSON SCHEMA:
{
  "hasIssues": boolean,
  "issues": [
    {
      "line": number,
      "severity": "critical" | "warning",
      "description": string,
      "recommendedFix": string
    }
  ]
}

const promptMessages = [
  { role: "system", content: "You classify support tickets into departments." },
  // Example 1
  { role: "user", content: "I cannot log in with my password." },
  { role: "assistant", content: JSON.stringify({ category: "auth", priority: "high" }) },
  // Example 2
  { role: "user", content: "Can I get a copy of my invoice from March?" },
  { role: "assistant", content: JSON.stringify({ category: "billing", priority: "medium" }) },
  // Real User Request
  { role: "user", content: "My account was charged twice today." },
];

import OpenAI from "openai";

const openai = new OpenAI();

async function extractUserData(text: string) {
  const response = await openai.chat.completions.create({
    model: "gpt-4o-mini",
    response_format: { type: "json_object" },
    messages: [
      {
        role: "system",
        content: "Extract user contact info as JSON with keys: 'name', 'email', 'phone'.",
      },
      { role: "user", content: text },
    ],
  });

  const parsed = JSON.parse(response.choices[0].message.content || "{}");
  return parsed;
}

import { z } from "zod";

// Define strict schema contract for model output
const OutputSchema = z.object({
  topic: z.string(),
  summary: z.string().max(200),
  confidenceScore: z.number().min(0).max(1),
  suggestedTags: z.array(z.string()).min(1).max(5),
});

export async function parseModelResponse(rawText: string) {
  try {
    const json = JSON.parse(rawText);
    const validated = OutputSchema.parse(json);
    return { success: true, data: validated };
  } catch (error) {
    console.error("Model generated invalid schema:", error);
    return { success: false, error: "Malformed AI response" };
  }
}