Claude Codeand Cursor, and the difference in velocity is staggering.
Instead of spending forty minutes debugging a specific middleware configuration or wrestling with a weird CSS quirk, I find myself describing the "vibe" or the intended behavior of the feature. I tell the LLM agent exactly what the user experience should feel like, and I let the agent handle the heavy lifting of the implementation details.
The shift from syntax to intent #
In a traditional workflow, your brain is constantly context-switching between high-level logic and low-level syntax. You think about a database schema, but then you get stuck on whether a specific library requires an async/await wrapper or a callback.
With an agentic AI workflow, that friction disappears. Here is how I've structured my current setup to maximize this:
-
High-level architectural prompting: I start by defining the system constraints. I don't just say "make a login page." I provide a prompt that defines the tech stack and the security expectations.
-
Iterative refinement: I use a "vibe-check" loop. I run the code, see the result, and if it feels clunky, I give a qualitative instruction like "make the transitions smoother" or "make the error handling more robust."
-
Agentic debugging: Instead of reading stack traces, I pipe the error directly into my terminal agent.
A practical example of the agentic loop #
When I'm working on a new feature, I don't write the boilerplate. I use a command-line approach that looks something like this:
claude dev "Create a new Express route for user profile updates.
Ensure it uses Zod for validation and integrates with our existing
PostgreSQL connection pool. Follow the pattern in /src/routes/auth.ts"
The agent doesn't just dump code; it reads my existing files, understands my architectural patterns, and implements the new code so it actually fits the project. This is the core of what people mean by agentic engineering—the AI isn't just a autocomplete tool; it's a collaborator that understands the context of the entire repository.
The reality check #
Is this going to replace developers? Not really. It's just raising the floor. If you don't understand how a database works or what a race condition is, you won't be able to "vibe code" your way out of a production meltdown. You still need to be the architect. The person who knows how to direct the agent is much more valuable than the person who can just write a loop in Python.
We are essentially moving toward a "Natural Language Compiler" era. The goal is to reduce the distance between a thought and a functioning piece of software. If you can describe it clearly, you can build it.
Next Coding agents are currently hitting a massive wall when it comes →
a library of Claude prompt techniques, with plenty of directly applicable cases.