Prompt Engineering for Code

The quality of your vibecoding output depends heavily on how you communicate with the AI. This guide covers proven techniques for writing better prompts.

The Anatomy of a Good Prompt

A great coding prompt has four parts:

1. Context — What exists already, what framework you're using
2. Goal — What you want to achieve
3. Constraints — Requirements, limitations, or preferences
4. Format — How you want the output structured

Example

Context: I have a Next.js app with Tailwind CSS and a
PostgreSQL database using Prisma ORM.

Goal: Create an API endpoint that returns a paginated
list of blog posts.

Constraints:
- Use App Router API routes
- Support page and limit query parameters
- Return total count in the response
- Handle errors gracefully

Format: Give me the API route file and the Prisma query
separately so I can review each.

Technique 1: Be Specific About Technology

Instead of: "Build a website"

Say: "Build a static website using HTML, CSS, and vanilla JavaScript. No frameworks."

Specifying your tech stack prevents the AI from making assumptions that don't match your project.

Technique 2: Provide Examples

Show the AI what you want:

Create a function that formats dates like these examples:
- "2026-03-07" -> "March 7, 2026"
- "2026-01-15" -> "January 15, 2026"
- "2025-12-25" -> "December 25, 2025"

Examples eliminate ambiguity better than descriptions.

Technique 3: Ask for Explanations

Add "explain your choices" or "add comments explaining the logic" to your prompts. This helps you:

• Learn from the generated code
• Catch mistakes or questionable decisions
• Build your own understanding

Technique 4: Constrain the Output

Without constraints, AI tends to over-engineer. Be explicit:

• "Keep it simple — no unnecessary abstractions"
• "Use only standard library features, no external packages"
• "This is a prototype, optimize for readability over performance"

Technique 5: Iterate with Context

When refining code, reference what you already have:

The login function you wrote works, but I need these changes:
1. Add rate limiting (max 5 attempts per minute)
2. Log failed attempts to the console
3. Return a more descriptive error message
Keep everything else the same.

This prevents the AI from rewriting code that already works.

Common Mistakes

• Too vague: "Make it better" — Better how? Be specific.
• Too much at once: Asking for an entire app in one prompt leads to generic, shallow code.
• No context: The AI doesn't know your project structure unless you tell it.
• Ignoring output: Not reviewing generated code leads to bugs and security issues.

Practice Exercise

Try this: Pick a simple feature you want to build and write three versions of the prompt — vague, moderate, and detailed. Compare the outputs. You'll see the difference specificity makes.