Artificial intelligence has transformed software development. More and more teams are adopting tools like GitHub Copilot and ChatGPT to speed up coding, documentation, and debugging. But there's a huge gap between promise and practicality: without a clear strategy, using AI can even be counterproductive.
In this article, I explain how to apply the "vibe-coding" approach, a structured methodology for working with AI from the top down, maintaining control, reducing errors, and maximizing impact. And best of all: with real-life examples, effective prompts, and techniques you can apply today.
Vibe-coding: Start with context, not code
One of the common pitfalls when using code generation tools is jumping straight into requesting features or snippets. This often results in confusion, wasted time, and inconsistent results.
The top-down approach proposes the opposite: starting with requirements , moving on to system design , then to code scaffolding , and finally to functional block generation . This allows you to control the process and correct conceptual errors before they spread.
Real-life example: indexing Cassandra in ElasticSearch
Initial prompt (summary):
I'm a PM, you're a software architect. We're going to design an c level contact list app to index a large Cassandra table in ElasticSearch. Help me build a robust, data-loss-free system.
This starts a guided Q&A session, where the AI explores:
Index use case (search, analytics...)
Data volume and insertion rate
Dataset type (insert-only, updatable, etc.)
In the end, the AI summarizes a proposed architecture:
Cassandra → Kafka → ElasticSearch (bulk writes)
Fault tolerance with DLQ queue
Multi-region replication
Corporate security and recovery
And it offers a complete technical specification.
Scaffolding: structure before content
Once the design is validated, you can ask the AI to generate the folder, class, and function schema:
Prompt:
Create a PHP app schema that reads from Kafka and multicasts to PostgreSQL, REST API, and SMTP.
Expected output:
This allows architectural errors or improvements to be detected before a single functional line is written.