AI Engineering in Practice: From Vibe Coding to Enterprise-Grade Development

Introduction: The Limitations of Vibe Coding and the Real Needs of Enterprise Development

As AI coding tools sweep through the developer community, "Vibe Coding" has become a buzzword. The concept was first coined by Andrej Karpathy in early 2025 — the former Tesla AI Director and OpenAI co-founder described a radically new way of programming on social media: developers fully immerse themselves in the "vibe," describe their intent in natural language, let AI generate all the code, and don't even bother reading the output. The idea went viral, with countless influencers claiming that Claude Code or Cursor could "replace programmers." But the reality? All they built were toy projects like ring lights and Pomodoro timers. Real enterprise projects involve complex business logic, distributed architectures, and high-concurrency scenarios — far beyond what simple prompts can handle.

A Chinese tech educator known as Teacher Zhuge presented a comprehensive, hours-long hands-on course systematically demonstrating the full progression from Vibe Coding to AI Engineering — covering practical usage of mainstream tools like Claude Code, Codex, and Cursor, as well as the engineering-oriented programming frameworks being adopted within enterprises. This article distills the core insights to help developers build the right mental model for AI-assisted programming.

Three AI Programming Modes: From Toys to Enterprise-Grade

Mode 1: Vibe Coding — Fast Prototyping, But Stops at the Demo

The core idea of Vibe Coding is simple: describe the requirements in your head using natural language and let AI coding tools generate the code. This approach can indeed scaffold a basic e-commerce project in minutes, making it highly effective for product managers validating ideas or developers quickly building prototypes.

But the problems are equally obvious — as the project scales, code quality deteriorates rapidly. The generated code is often "dead-weight code" (unmaintainable code), and when bugs appear in production, non-technical users simply cannot effectively direct AI tools to fix them, easily trapping the project in a death spiral. Karpathy himself acknowledged that Vibe Coding is better suited for one-off prototype validation rather than production-grade development. Critics point out that code generated this way lacks maintainability and extensibility, making it virtually unusable in real team collaboration and long-term iteration scenarios.

Mode 2: Plan Mode — Introducing Structured Thinking

Taking it one step further, both Claude Code and Codex offer a Plan Mode. In this mode, the AI doesn't jump straight into writing code. Instead, it first generates a detailed development plan — including technology choices, module breakdown, implementation steps, and more — and only proceeds step by step after the developer confirms.

The technical principle behind Plan Mode is closely related to the "Reasoning-Acting" paradigm of AI Agents. In traditional single-turn code generation, the large model directly maps natural language requirements to code output, which easily leads to context loss and logical inconsistencies. Plan Mode introduces the Chain-of-Thought concept, having the model first perform Task Decomposition to generate a structured execution plan before implementing it step by step. This approach essentially combines the "design before code" waterfall model philosophy from software engineering with AI's reasoning capabilities, significantly improving the quality of complex task completion.

This approach is far more reliable than pure Vibe Coding because developers can spot potential issues during the planning phase and make adjustments. However, for truly complex enterprise projects, it still isn't systematic enough.

Mode 3: AI Engineering — The Right Approach for Enterprise Development

True enterprise-grade AI programming requires a complete engineering workflow. The course highlights a Claude Code plugin called SuperPower, which is essentially an Agent Scale (agentic workflow) for AI engineering — containing over a dozen or even dozens of Skills that cover the entire pipeline from requirements analysis, architecture design, code development, and testing to deployment.

Agent Scale (agentic workflow) is an important architectural pattern in current AI application development. Unlike the simple "prompt-response" model, agentic workflows decompose complex tasks into multiple specialized Agents, each responsible for specific subtasks, collaborating through predefined workflows. For example, the SuperPower plugin may include a Requirements Analysis Agent, Architecture Design Agent, Code Generation Agent, Code Review Agent, Test Generation Agent, and more. This architecture borrows from microservices design principles — each Agent has clear responsibility boundaries and input/output specifications, coordinated through an Orchestration Layer, enabling complex task processing capabilities far beyond what a single AI model can achieve.

The core philosophy of this system is SDD (Specification-Driven Development), similar in concept to earlier tools like Spy Code and Spy Kit, but far more mature and systematic. SDD's intellectual roots trace back to several classic software engineering methodologies — it shares lineage with TDD (Test-Driven Development) and BDD (Behavior-Driven Development), but the key difference is that SDD treats "specification documents" as the Single Source of Truth for the entire development process. In the AI programming context, this means developers first write detailed technical specifications (including API interface definitions, data models, business rules, etc.), and AI tools generate code strictly according to those specifications. This approach effectively addresses the common "requirement drift" problem in AI programming — the phenomenon where AI gradually deviates from the original requirements across multiple conversation turns. A growing number of small and medium-sized companies are already adopting this workflow in real projects.

Tool Selection: Claude Code vs Codex vs Domestic Alternatives

Claude Code: Most Capable, But Highest Barrier to Entry

From a pure technical capability standpoint, Claude Code remains the most widely used and most powerful AI coding tool among professional programmers. It features a highly refined internal AI engineering system with extensive optimizations for professional programming.

However, the practical issue is that Anthropic's policies toward Chinese users are unfriendly — Claude's latest models have been restricted to US-only access, account bans are frequent, and recovery is extremely difficult. Teacher Zhuge candidly shared that his own Claude account has been banned multiple times, and he now hesitates to use it.

Codex: Catching Up Fast, Increasingly Practical

OpenAI's Codex previously lagged significantly behind Claude Code, but with the release of GPT's latest versions and continuous internal optimization, the gap is closing rapidly. The course demonstrated development using both Codex's desktop client and CLI (command-line interface), showcasing multiple development workflows commonly used in enterprises.

Domestic LLMs: JAM Leads, Each with Unique Value

For backend model selection, the course provided a practical assessment of domestic Chinese LLMs' programming capabilities:

• Tier 1: JAM (StepFun) — Best overall experience; despite being a young company, its valuation is approaching Xiaomi's
• Best Value: DeepSeek — Solid programming capabilities at extremely low prices
• Tier 2: MiniMax, KIMI, MIMO (Xiaomi), Tongyi (Alibaba), Hunyuan (Tencent) — Each has its strengths and all are viable options

The rapid improvement of domestic LLMs in programming capabilities is driven by multiple technical factors. First, in terms of training data, open-source code repositories on GitHub, Q&A data from Stack Overflow, and various technical documentation provide rich code training corpora. Second, the application of reinforcement learning techniques (RLHF/RLAIF) enables models to optimize based on code correctness, efficiency, and readability. StepFun's standout performance in programming is closely tied to its specialized optimizations in code understanding and generation. DeepSeek achieves exceptional cost-effectiveness through its MoE (Mixture of Experts) architecture — maintaining strong programming capabilities while dramatically reducing inference costs. It's worth noting that programming capability assessments typically reference standardized benchmarks like HumanEval, MBPP, and SWE-bench, but real-world engineering performance often differs from benchmark results.

All these models can be accessed through aggregation platforms like OpenRouter for use with Claude Code, avoiding the account ban risks of using Claude's API directly. OpenRouter is an LLM API aggregation platform whose business model resembles API gateways in cloud computing or payment aggregators in fintech. Technically, OpenRouter provides a unified API interface — developers only need to integrate once to call dozens of different vendors' LLMs (including Claude, GPT, Llama, Mistral, and various domestic models). The platform profits by charging a small markup on each API call (typically 5%-15% above the original price). For Chinese developers, accessing models like Claude through OpenRouter effectively mitigates the account risks associated with using official APIs directly.

Development Environment and AI Programming Toolchain Setup

VS Code: The IDE of Choice in the AI Era

The course explicitly recommends VS Code as the primary development tool, going so far as to say that if traditional IDEs like IntelliJ don't undergo major transformation, "the death knell will ring in two to three years." The reasoning is straightforward: VS Code's plugin ecosystem offers the most comprehensive support for AI programming tools, and Cursor itself is built on top of VS Code.

VS Code's dominant position in the AI programming era stems from its open plugin architecture and vibrant developer ecosystem. Built on the Electron framework and adopting open protocols like Language Server Protocol (LSP), VS Code enables deep integration of third-party tools into the editor. By contrast, JetBrains IDEs (such as IntelliJ IDEA), while offering a more mature traditional development experience, have a relatively closed plugin system and heavier runtime environment, causing noticeably slower adaptation of AI programming tools. Next-generation AI coding tools like Cursor and Windsurf have almost universally chosen to build on VS Code, further cementing its ecosystem advantage in AI programming. Mainstream AI coding plugins like GitHub Copilot, Cline, and Continue also prioritize VS Code support.

The specific development environment configuration is as follows:

• Base IDE: VS Code
• AI Coding Plugins: Claude Code plugin (dialog mode) + SuperPower plugin
• CLI Tools: Claude Code CLI + Codex CLI
• Desktop Client: Codex Desktop
• Model Backend: Domestic models like JAM accessed via OpenRouter

The Core Formula for AI Programming

Teacher Zhuge summarized a key insight: AI Programming Capability = Backend Model (primary factor) + AI Programming Tool (secondary factor). Tools are merely vehicles — what truly determines code quality is the underlying LLM's capability. This is why even when using the same tool, connecting different models produces dramatically different results.

Business Insights in the AI Space: Who's Actually Making Money

The course also shared an interesting business observation: the real money in AI right now isn't coming from consumer-facing applications (products like Doubao and Yuanbao are all losing money), but from three types of businesses:

1. Selling Compute: Chip, semiconductor, and memory hardware manufacturers — their stock prices have skyrocketed
2. Selling Tokens: API services from major players like Zhipu and Tencent
3. Aggregation Platforms: Model aggregation sites like OpenRouter — teams of a dozen people generating hundreds of millions in annual revenue

This is also why the course chose to build an "AI Model Aggregation Platform" as its second hands-on project — it's not just a technical exercise but a genuinely viable business model.

Enterprise Practice: Alibaba's Internal AI Engineering System

The course also mentioned the AI engineering programming system being rolled out internally at Alibaba — a complete "self-contained evolution" system. While confidentiality constraints prevent full hands-on practice, its core philosophy is to elevate AI programming from individual tool usage to team-level and organization-level engineering standards, representing the future direction of AI programming in enterprises.

Conclusion: Engineering Thinking Is the Core Competitive Advantage in AI Programming

AI programming is rapidly moving from the novelty phase of "Vibe Coding" into the deep waters of engineering and professionalization. For developers, the key isn't learning to use a specific tool to generate code, but understanding the complete AI engineering workflow and mastering the systematic methodology from requirements analysis to architecture design to code implementation.

Tools will keep evolving, but engineering thinking is the true core competitive advantage. Whether you choose Claude Code, Codex, or domestic LLM solutions, embracing the SDD (Specification-Driven Development) philosophy and establishing structured AI programming workflows is what will truly convert AI programming capabilities into enterprise-grade productivity.

Key Takeaways