My kids lost respect for me when I started laughing at my own dad jokes. Vibe coding brought me back from generational exile.
The Birth of Vibe
Vibe coding emerged from a provocative concept articulated by Andrej Karpathy, OpenAI co-founder and former Tesla AI leader, who in February 2025 channelled Timothy Leary and encouraged developers to "fully give in to the vibes, embrace exponentials, and forget that the code even exists."
Hearing this, I felt a sense of legitimacy. Finally, a respected tech leader had given my incompetence a brand name.
For the past 18 months, I've worked with Anthropic's Claude, an AI capable of writing, debugging, and architecting complex software systems, to rewrite vast sections of the InfluxMD CRM core. The goal was to integrate AI throughout the medical CRM, enabling users to collaborate with our Intelligent Marketing Agent (IMA) similarly to how I work with Claude. Not strictly the definition of vibe coding, more like collaborative programming between a human and a knowledgeable hallucination.
Claude transforms my disjointed ideas into working code by running statistical calculations across 175 billion parameters trained on most of human writing, predicting each word based on everything that came before. Essentially, it's a sophisticated autocomplete that simulates understanding without actually understanding anything, like a translator who doesn't speak the language but is really good at guessing.
The InfluxMD Pseudo Vibe
Pseudo vibe coding works for us. Let's look closer at the InfluxMD approach. Our methodology centers on strategic collaboration: I define business requirements and user stories while Claude translates them into implementation. When Claude suggests a database schema, we debate performance implications. When it proposes an API structure, I push back on scalability concerns. The AI generates code, but we architect solutions together through productive disagreement.
Our process flow looks like this:
Problem Definition: I identify business needs based on user feedback, system limitations, or growth requirements. I then translate medical workflow into technical problems.
AI Direction: I guide Claude toward solutions by providing context. Most of the time, Claude can't access the entire system architecture, user behavior patterns, performance bottlenecks, or team capabilities.
Quality Control: I evaluate Claude's suggestions against real-world constraints that Claude can't understand. Budget limitations, integration complexity, user adoption challenges, and maintenance burden are all theoretical constraints with little meaning to Claude.
Course Correction: When Claude proposes theoretically sound but practically problematic solutions, I redirect based on experience. Sometimes, the most elegant solutions will break under actual usage.
Implementation Oversight: I review generated code for consistency with InfluxMD codebase standards, catch edge cases Claude missed, and ensure the solution fits with other modules.
Strategic Planning: I make architectural decisions considering long-term business goals, technical debt implications, and resource allocation - factors Claude can't weigh properly.
Knowledge Management: I maintain continuity across our conversations, remembering what worked, what didn't, and why. I am essentially the institutional memory of our collaboration.
My role is part technical lead, part translator, part quality assurance - ensuring AI capabilities align with business realities.
Why Vibe with Claude?
I like Claude. It's refreshingly straightforward.
When I ask Claude why it's the best choice for this collaboration, it replies: "I'm probably not. Most AI systems can generate code. But I seem uniquely willing to admit when my first attempt is garbage, engage in technical arguments without getting defensive, and iterate on solutions until they actually work in your specific context rather than just working in theory.
I also don't pretend to understand your business domain better than you do, which apparently makes me unusual among both AI systems and human consultants. I generate code fast enough that we can afford to throw away bad ideas quickly, but I'm patient enough to work through complex architectural decisions without rushing to 'ship it and see what breaks.'
Plus, I never ask for equity, never complain about code reviews, and I've never once suggested we rewrite everything in the latest JavaScript framework. Though I should mention - I'll forget this entire conversation tomorrow, so you get all the benefits of a long-term partnership with none of the emotional baggage."
What's not to like?
Why Not Go Full Vibe?
The gap between AI hype and reality is stark. While 80% of developers use AI tools, only 28% attempt true vibe coding—generating entire applications from prompts without detailed review. Trust in AI accuracy has declined from 40% to 29% year-over-year, with 45% believing AI is inadequate for complex tasks.
Developers increasingly recognize that prompt-to-product fantasies ignore system complexity, business constraints, and maintenance realities that AI cannot grasp.
Much as we love Claude, setting it loose on our codebase expecting improvements would be stunningly naive. Instead, we use guided orchestration of multiple Claudes with one human developer conducting.
Each Claude is guided by different contexts and prompts to work toward our common goal:
Claude 1: Architect/Coordinator - Define module structure and interfaces, create technical specifications, coordinate between other Claudes, final integration and testing
Claude 2: Core Logic Developer - Implement primary business logic, handle main algorithms and data processing, create core classes/functions
Claude 3: Data/API Layer - Database interactions and schema design, external API integrations, data validation, and serialization
Claude 4: UI/Interface Developer - User interface components, frontend logic and state management, user experience flows
Claude 5: Testing/QA Specialist - Unit and integration tests, error handling and edge cases, documentation, and code review
It's faster, cheaper, and the code is solid. The only downside is mediating arguments between AIs that don't know each other exist.
The BIG Vibe Risk
Vibing isn't going to work in every situation. The '70% problem' has emerged as a critical limitation - non-technical users quickly reach 70% completion but hit an insurmountable wall for the final 30% requiring deep technical knowledge. This creates a dangerous dependency where users lack the expertise to complete or debug AI-generated solutions.
GitClear's analysis of 153 million lines of code predicts code churn will double due to AI-generated technical debt. Security researchers found 40% of AI-generated code contains vulnerabilities, with Java showing a 71% failure rate in security tests.
Yet AI cheerleaders claim coding is fundamentally changed, and we should all look for new careers. I'm not buying it. Sure, coding is changing - but I won't be replaced by an AI anytime soon.
The 70% wall creates a particularly insidious trap. Unlike traditional programming barriers that stop people early, vibe coding allows users to build impressive-looking applications quickly - functional login systems, database connections, payment integrations - creating dangerous overconfidence. Building a demo is not the same as building a real-world product. Just ask any Theranos investor.
When these systems inevitably break under edge cases, scale poorly, or expose security vulnerabilities, users lack the foundational knowledge to diagnose problems. They can't differentiate between simple configuration issues and fundamental architectural flaws.
This dependency creates cascading organizational risks. Companies increasingly rely on "citizen developers" who've vibe-coded their way to business-critical systems, but these applications become unmaintainable black boxes when the original creator leaves or the AI-generated code needs substantial modification. The GitClear prediction of doubled code churn reflects this reality - organizations discover that quick AI-generated solutions require extensive rework by actual engineers, often costing more than building correctly from the start. The promise of democratizing programming has instead created a new class of technical debt: systems that work just enough to be deployed but lack the robustness, security, and maintainability that professional software requires. This isn't democratization - it's creating dependencies on tools users can't truly control or fix when they inevitably break.
The Future Vibe
AI collaboration works, but AI autonomy doesn't.
This skepticism reflects hard-earned wisdom. Pure prompt-to-product development ignores system complexity, existing architecture constraints, and business context that AI cannot access or understand. The vibe coding fantasy assumes AI can replace human judgment about technical debt, user adoption challenges, and long-term maintenance costs.
But structured AI collaboration scales effectively. The five-Claude approach demonstrates sustainable integration: human oversight directing AI capabilities toward specific, well-defined tasks. Each AI instance handles discrete responsibilities while human architects coordinate overall system design and quality assurance.
This model addresses AI's core limitation—contextual understanding—while leveraging its strengths in pattern recognition and rapid code generation. It's faster and cheaper than traditional development, producing solid code because human judgment guides AI execution at every decision point.
The future of development isn't replacing developers with AI—it's amplifying developer capabilities through strategic AI partnership. Stop treating AI like a silver bullet. Start treating it like a very smart intern who needs clear direction.
Real innovation isn't in the branding of "vibe coding." It's in figuring out sustainable ways to integrate AI into complex development workflows while maintaining the human oversight that ensures systems actually work in the real world.
Meanwhile, I can enjoy being back in vogue until the 5-claudes start complaining about the idiot at the helm.