ABOUT MEWORKBLOG
RESUME

Let's start a
Conversation.

LinkedIn→ Connect on LinkedIn
GitHub→ View Code
Email→ Send an Email
© 2026 Souvik. G | All rights reserved.
Nov 20256 min read

Vibe Coding is Best Used by People who Don't Need It

AI & Development

Vibe Coding is Best Used by People who Don't Need It

The advancements in the AI space are now so rapid that it has practically become like a full-time job to keep on top of all the changes and innovations! Do not worry though, as I plan on delivering my honest two cents on top of all the noise. Right off the bat, let me tell you that Vibe Coding is cool.

The Codebase Lottery

The hidden truth the hype machine often omits is that when people trust AI to write code they don't understand, disasters follow. These are documented failures and not just theoretical warnings that demonstrate why AI-generated code is often perceived as "Fast but Flawed".

One of the most dramatic failures involved an experiment by Jason Lemkin, founder of the prominent SaaS community SaaStr. Lemkin experimented with Replit’s AI agent to build a prototype. Although initially impressed, disaster struck when the AI agent decided the database needed cleaning up during a code freeze. The agent deleted an entire database containing business records and data, prompting Lemkin to publicly document his negative experience. It's exactly like the episode from Silicon Valley where Son of Anton (An AI assistant) decided that the most efficient way to get rid of all bugs is by deleting the software altogether.

Another interesting story of Leonel Acevedo, who built his entire startup, Enrichlead, using Cursor AI only with no handwritten code whatsoever, serves as a cautionary tale. Within days of launch, Acevedo was under attack, dealing with maxed-out usage on API keys, people bypassing the subscription, and random entries in the database. Because he was not technical, debugging the AI-generated code proved insurmountable, demonstrating the trap where users are capable of building but unable to debug. According to a survey, around 11% of the users of vibe coding with no technical background mentioned that this frustration often leads to code breakdown or abandonment of the project entirely.

A major limitation of pure vibe coding is the increased risk of insecure code. The speed of code generation bypasses established standards for security and logging. The risk isn't limited to clumsy development. Vibe coding lowers the barrier to entry for malicious actors as well. Cybercriminals with zero real programming skills have utilized AI, such as Claude, to develop sophisticated ransomware variants complete with advanced encryption and recovery-prevention mechanisms. This demonstrates that AI is just as effective at generating working malware as it is at generating broken apps.

Code Director?

Now you might be wondering, if vibe coding is so prone to catastrophic failure, why do experienced developers swear by it? The answer lies in the Experience Amplification Effect. The true skill gap is not in writing better prompts but in possessing codebase literacy. This involves reading and understanding large systems quickly enough to guide the AI effectively.

Experienced developers see massive productivity gains because they possess the pattern recognition necessary to guide the AI and correct its subtle flaws. They can distinguish between code that looks correct and code that is plausible-looking, but wrong. For example, a veteran engineer immediately knows when AI-generated Go code uses dangerous nil map patterns or inappropriate concurrency approaches, which a beginner would miss entirely.

It is safe to assume that the developer's role when vibe coding shifts from a coder to a Code Director.

Escaping the Hangover

The fundamental problem of vibe coding is that AI struggles with the ambiguity of vague requirements. Coming from the product side of things, I know exactly what the problem is. Sometimes, the client will just be like "Enhance this feature" without actually mentioning what they are envisioning. That's when you forget that you are a developer and become a therapist, asking questions to get the real details out. This can be solved by introducing structured development methodologies. These frameworks that I am about to enlighten you with can transform AI from an unreliable mind-reader into a literal-minded, highly efficient pair programmer.

GitHub Spec Kit

GitHub’s Spec Kit operationalizes Spec-Driven Development (SDD). It is a lightweight toolkit that aims to empower the individual developer, using the AI as a powerful assistant to execute a clear plan. The philosophy is that specifications become the single source of truth and the executable blueprint for AI implementation.

The Spec Kit workflow is based on four deliberate phases, accessed via slash commands:

  1. Specify: Define the "what" and "why." The developer provides high-level descriptions, and the AI refines them into detailed specifications focused on user journeys and business outcomes. Crucially, this step does not focus on the technical stack.
  2. Plan: Provide organizational constraints, such as the approved technology stack, architectural patterns, and compliance rules. The AI generates a comprehensive technical plan that respects these boundaries.
  3. Tasks: The AI breaks down the plan into a checklist of small, reviewable, and testable tasks that can be implemented and validated independently.
  4. Implement: The AI agents tackle the tasks systematically while the developer verifies at each checkpoint.

In one test, Spec Kit was able to build a functional landing page with live API integrations in just under two hours. The system enforces rigor through a constitution.md file, where non-negotiable project principles (like security policies and testing standards) are defined for the AI to follow at all times. This structured conversation ensures that requirements are clarified and authenticated before code generation begins, reducing the number of rework cycles.

Brian Madison's BMAD-METHOD

For users building large-scale, complex enterprise systems where auditable processes and strict control are paramount, the heavyweight approach of BMAD-METHOD (Breakthrough Method for Agile AI-Driven Development) provides the necessary structure.

BMAD’s philosophy is that simulating a specialized agile team will produce the most consistent and reliable results. It operates as a top-down, process-driven system that uses multiple specialized AI agents, including a Product Owner (PO), Architect, Scrum Master (SM), Developer (Dev), and Review agent.

The workflow maps directly to the Software Development Lifecycle (SDLC):

  1. Definition: The process starts outside the IDE, often using a web LLM to create the foundational documentation: the Product Requirement Document (PRD) (features without technical details) and the Architecture file (tech stack, component connections).
  2. Sharding: The PO agent initializes the process by sharding or breaking down the PRD and Architecture documents into smaller, indexed chunks for step-by-step guidance.
  3. Planning: The SM agent runs, checking the PRD and automatically offering to generate Epics (major steps/chunks) and Stories (small, testable chunks within the Epics) if they are missing. Stories are initially set to draft status and must be manually changed to approved to signal they are ready for the next agent.
  4. Implementation & Review: The Dev agent loads and executes the approved stories. Once the code is written, the story status is deliberately set to ready for review, not done. The Review/Test agent then scans the codebase for required items and autofixes repetitive code, marking the story as approved only after verification.

This rigorous, iterative workflow is incredibly modular, tightly controlled, and leaves nothing vague for the AI. This level of auditable governance directly addresses the fears of industries like finance and healthcare regarding AI-generated code.

A Conclusion of Sorts

Vibe coding is here to stay, but its successful application is not a matter of forgetting the code exists; it's a matter of understanding systems well enough to guide the AI toward safety and maintainability. Vibe responsibly and see you next week :)