Prompt Engineering for Code

The art of formulating your requests well

Prompt engineering is the most important skill in Vibe Coding. A good prompt produces functional code on the first try. A bad prompt generates off-topic or buggy code.

Fundamental principles

1. Be specific

Bad prompt:

Make me a form

Good prompt:

Create a registration form in Angular with email, password, and
password confirmation fields. Use Reactive Forms with validation:
valid email, password min 8 characters, confirmation matching
the password. Display errors below each field.

2. Provide context

The AI doesn't know your project. Provide it with:

The framework being used (Angular, React, NestJS...)
The project's conventions (file structure, naming)
The available dependencies (TypeORM, Tailwind, etc.)
The purpose of the feature in the overall context

3. Break down complex tasks

graph TD
    A[Complex task] --> B[Sub-task 1]
    A --> C[Sub-task 2]
    A --> D[Sub-task 3]
    B --> E[Prompt 1]
    C --> F[Prompt 2]
    D --> G[Prompt 3]
    E --> H[Integrated code]
    F --> H
    G --> H

Rather than asking "build me a complete e-commerce app", break it down:

Create the data model for products
Create the CRUD API for products
Create the catalog page with filters
Add the cart with quantity management

4. Iterate progressively

Don't aim for perfection on the first try. Proceed by iterations:

V1: minimal working version
V2: adding style and UX
V3: error handling and edge cases
V4: optimizations and refactoring

Effective prompt patterns

The "role + context + task" prompt

You are a senior NestJS developer.
I'm working on a REST API for an educational platform.
Create a service that manages user progress in programs:
- Mark a chapter as read
- Calculate the progress percentage
- Use TypeORM with a UserProgress entity

The "example + adaptation" prompt

Here is my existing service for programs:
[paste the code]

Create a similar service for quizzes, following the same
conventions and patterns.

The "constraint" prompt

Create an Angular component to display a list of program cards.
Constraints:
- No external library, only Tailwind CSS
- Standalone component
- Responsive: 1 column mobile, 2 tablet, 3 desktop
- Each card displays: title, description, progress, image

Common mistakes

Too vague

"Improve this code" -> The AI doesn't know what you mean by "improve".

Too long

A 500-line prompt will overwhelm the AI. Stay concise and focused.

No verification

Copy-pasting AI code without reading it is dangerous. Always review.

Ignoring errors

When the AI generates code that doesn't compile, don't resubmit the same prompt. Share the exact error with the AI.

Summary

Prompt engineering for code rests on four pillars: specificity, context, decomposition, and iteration. The more you practice, the more effective your prompts become and the higher the quality of generated code.

Vibe Coding : Coder avec l'IA

Chapitres