AI & Agents
Vibe Coding Is Overrated Unless You Can Judge the Code
AI writes code fast. It cannot tell you whether that code should exist. The skill everyone is skipping is engineering judgment, and here is the workflow I use to apply it before a single prompt.
vibe codingvibe coding production readyAI generated code review
Before I give a builder agent a single prompt, I spend my time arguing with a different AI about the database. Not about features. About what happens to the schema at 100,000 rows, and which decision I will hate most in six months.
I have 5+ years behind me as a full-stack developer and about 10 as a Senior PM, and I vibe code a lot. So I will say the unpopular part plainly: vibe coding is overrated for anyone who cannot judge the code that comes back. The AI is fast. It has no opinion on whether the thing it just built should exist, or whether it was built the right way.
That judgment is the skill most people are skipping.
What does AI actually build when you just prompt and ship?
Sit a non-technical founder in front of a coding agent and say “build me this product.” You get a clean React frontend, a hosted database wired in, routes, components, some basic auth. It looks like a product. In a demo it behaves like one.
Then look at what nobody asked:
- Database design. What happens to the schema when the user table hits 100,000 rows? Did anyone think about indexing, or about what a badly designed foreign key does to query time six months from now?
- Security and auth. Is session handling actually secure? Are API keys sitting in the frontend? Is there rate limiting on the login endpoint, or is the product one script away from credential stuffing?
- Backend shape. Monolith or not? What happens when one service becomes the bottleneck? And who made that call, you or the AI defaulting to the simplest pattern it could generate?
- Infrastructure. Where does it live, what does deployment look like, and is the rollback plan a plan or just panic?
- Testing. Real tests. Unit, integration, end-to-end for the edge case your happy-path demo never touched. Clicking around and deciding it seemed fine does not count.
Left alone, the AI answers none of these. It builds what you asked for, and what you asked for is rarely what you needed.
Is vibe coding a revolution or dangerous nonsense?
The loud camps are two. One says anyone can build anything and engineers are optional. The other says AI code is insecure, unmaintainable junk.
I think both are arguing about the wrong variable. The tool is fine. What breaks things is vibe coding without context. A prompt is only as good as the person writing it, and that person only knows what to ask for if they understand what they are building, architecture included, not just features.
When I sit down to vibe code I play two roles at once, the PM describing the product and the engineer in the room asking uncomfortable questions. What is the expected load on this service? What does failure look like and how do we recover? Where can we cut corners on purpose, and where can we not? Those are engineering questions. If nobody in the room asks them, the AI will not volunteer them.
How do I vibe code something I would actually ship?
My workflow has three steps, and the first two happen before any code exists.
Step 1: Architect before you prompt
I open a session with the AI as a thought partner, not a code generator. We work through the system design: infrastructure, monolith versus microservices, relational versus document store, serverless versus containers. We define the data model and map the API contracts. We list the edge cases that will hurt later.
I push back on its suggestions the way I would push back on an engineer in a design review. “What breaks here?” “What are we assuming?” Nothing gets built until I am satisfied with the architecture.
Step 2: Ask for prompts, not code
Most people skip this one, and I think it is the step that makes the rest work. I ask the AI to write the prompts for me, carrying the full spec: architectural constraints, performance requirements, security considerations, testing strategy, deployment approach.
Compare two briefs.
| Vague brief | Specified brief |
|---|---|
| ”Build a user authentication system." | "Build a JWT-based authentication system with refresh token rotation, rate limiting on the /auth/login endpoint set to 5 attempts per 15 minutes, secure httpOnly cookie storage, and error handling that does not leak user enumeration data.” |
| Works in a demo. | Something you can actually ship. |
The model is the same in both rows. The brief is different. Writing the second one takes an engineering background, or at least enough of one to know what belongs in it.
Step 3: Let the builder agent build
Only now does a builder agent get involved, and it is building with intent. The spec reads like it came from a senior engineer because the prompts were designed by someone thinking like one. When I review the codebase afterward, the decisions are traceable. I can see architectural thinking in it instead of whatever pattern was most convenient to generate.
How much engineering do you need to vibe code responsibly?
Less than you fear, and more than zero. You do not need to have shipped a distributed system or hand-written a Kubernetes config. You need enough context to ask the right questions:
- A database schema is a decision, not a detail.
- Auth is not a checkbox.
- “Works on my machine” and “works in production” are two different sentences.
- Infrastructure costs money, and “scale later” is a plan with consequences.
Without that, you are vibe coding a prototype and calling it a product. That matters to your users, to anyone you pitch, and to the engineer you eventually hire to repair what you built.
If your reason for vibe coding is revenue rather than craft, I wrote separately about why everyone is vibe coding and nobody is getting paid. The two posts point at the same gap from different ends.
Did AI lower the bar for building software?
For starting, yes. Hugely. A single person can ship things now that used to need a team, and I do not expect that to reverse.
For shipping something good, I would argue the bar went up. Everyone can produce a demo, so the demo stopped being the differentiator. What separates one build from another is whether it holds up, and holding up needs judgment that prompts cannot replace. I cannot prove this with data. It is what I see in my own builds and in the products I look at.
So a software engineering background still earns its keep in the AI years. It lets you use AI as a force multiplier instead of a replacement for thinking.
A checklist before your next vibe coding session
- Write down the load you expect and what failure should look like.
- Decide the data model yourself, or argue it out with the AI until you agree.
- Ask the AI for specified prompts, with security, performance and testing written in.
- Mark which corners you are cutting on purpose.
- Review the output for decisions, not just for whether it runs.
The AI will build anything you tell it to, quickly and confidently, without once asking if it is the right thing. That part was always the human’s job.