Claude Code + Harness AI Engineering in Practice: From Methodology to Enterprise-Level Implementation

Introduction: The Shift from Traditional Coding to AI-Engineered Programming

A profound transformation is underway in software development. More and more companies are delegating the bulk of their coding work to AI, and the developer's role is shifting from "the person who writes code" to "the person who directs AI to write code." According to reports, leading AI companies like Anthropic and OpenAI already have over 90% of their code written by AI large language models.

Behind this trend lies the rapid evolution of AI programming tools. From GitHub Copilot launching the era of LLM-driven code completion in 2021, to the emergence of intelligent IDEs like Cursor and Windsurf, to Anthropic's release of Claude Code in 2025 pushing AI programming into the new paradigm of "Agentic Coding" — AI is no longer just passively responding to instructions but can proactively analyze codebases, plan implementation paths, and execute multi-step development tasks. This leap in capability makes it possible for AI to evolve from an "assistive tool" to a "programming collaborator."

However, in practice, many developers have encountered various problems after trying AI programming, and some have even abandoned it to return to "traditional coding" — the purely manual approach. Where does the problem lie? The key issue is the lack of a systematic AI engineering methodology for programming. This article, based on a practical course on Claude Code + Harness AI engineering shared by Teacher Zhuge on Bilibili, outlines the core approach to implementing AI programming in enterprise-level projects.

What Is Harness AI Engineering for Programming?

It's More Than Just Having AI Write Code

Many people's understanding of AI programming stops at "throw requirements at AI and let it generate code." But in real enterprise projects, this rough-and-ready approach often leads to uncontrollable code quality, chaotic architecture, lost context, and a host of other problems.

The core philosophy of Harness AI engineering is: integrating AI programming into an engineering framework, using standardized processes, clear division of responsibilities, and systematic management to make AI a reliable "programming collaborator" rather than an uncontrollable "code generator."

The word "Harness" in engineering means "to control and utilize," and this methodology's core ideas draw from mature practices in traditional software engineering. Looking back at the history of software engineering — from the waterfall model to agile development to DevOps — each paradigm shift involved not just tool updates but also a restructuring of processes and methodologies. AI programming faces the same challenge: if you only introduce tools without changing workflows, the results will inevitably fall short. The essence of the Harness methodology is treating AI as a "special member" of the team, establishing clear input specifications (Prompt templates), output standards (code quality requirements), and collaboration processes (task orchestration mechanisms) to maximize the efficiency of human-AI collaboration.

The Human Role Hasn't Disappeared

Teacher Zhuge particularly emphasizes one point: although 80%-90% of code can be completed by AI, this doesn't mean humans have nothing to do. Quite the opposite — humans take on higher-level responsibilities in AI-engineered programming:

• Architecture design and decision-making: The overall system architecture and technology selection still require human oversight
• Requirement decomposition and task orchestration: Breaking down complex business requirements into task units that AI can execute
• Code review and quality assurance: AI-generated code needs manual review and optimization
• Context management and knowledge accumulation: Maintaining project context information to ensure AI understands the business background

Among these, context management is one of the most critical technical challenges in AI programming. A large language model's "Context Window" determines how much information it can process at once — for Claude, the context window is approximately 200K tokens, equivalent to about 150,000 Chinese characters of text. While this capacity is already quite substantial, in large enterprise projects, codebases often span hundreds of thousands of lines, far exceeding the capacity of a single context window. This causes AI to "forget" previous context when handling complex projects, producing inconsistent code. Claude Code mitigates this through mechanisms like automatic project file indexing and intelligent retrieval of relevant code snippets, but developers still need to proactively manage context information through reasonable task decomposition and context documents (such as the CLAUDE.md configuration file) to ensure AI always operates within the correct business context.

Technology Stack Selection: A Pragmatic Approach with Claude Code and Domestic LLMs

Why Choose Claude Code?

Claude Code is an AI programming tool released by Anthropic that excels in code comprehension and generation. Unlike IDE-based AI programming tools such as Cursor, Claude Code uses a command-line interaction approach, making it better suited for handling complex multi-file edits and project-level code refactoring tasks. The course uses Claude Code combined with VS Code as the development environment, which is one of the mainstream AI programming tool combinations in the industry today.

The Viability of Domestic LLMs

One noteworthy practical choice in the course is that the backend model doesn't use Claude's native model but instead employs domestic Chinese models such as ByteDance's Doubao. The reason is practical — using Claude's model requires a VPN, which poses compliance risks in enterprise environments.

Domestic LLMs have made significant progress in code generation capabilities in recent years. ByteDance's Doubao series, Alibaba's Qwen (Tongyi Qianwen), DeepSeek's DeepSeek-Coder, and others have approached or even matched international top-tier performance on multiple code benchmarks (such as HumanEval and MBPP). For domestic enterprises, using Chinese LLMs not only solves network access and data compliance issues but also offers lower API call costs and faster response times. Claude Code, as a client-side tool, supports connecting to different backend models through API gateways. This decoupled architecture of "frontend tool + backend model" allows developers to flexibly switch underlying models without changing their workflow.

Real-world testing has shown that the combination of Claude Code + domestic LLMs can already achieve quite impressive results. Teacher Zhuge points out that if you use Claude's native model with the same Harness AI engineering methodology, the results would only be better. This provides an important message for domestic developers: even with network environment constraints, AI-engineered programming is entirely feasible.

Full-Process Implementation for Enterprise-Level E-Commerce Projects

Key Elements of System Building

Implementing AI-engineered programming in enterprise-level e-commerce projects requires building a complete framework. This isn't something that can be solved by simply introducing an AI tool — it requires systematic construction across the following dimensions:

1. Project specification documentation: Documenting coding standards, architectural constraints, and business rules as AI's "knowledge base"
2. Task granularity standardization: Defining the standard granularity of tasks AI can handle, preventing tasks from being too large and causing AI to "lose its way"
3. Prompt engineering templatization: Establishing standardized Prompt templates for different types of development tasks
4. Quality check automation: Building automated code inspection pipelines to ensure quality control of AI-generated code

Among these, Prompt Engineering is far more important in AI programming scenarios than in everyday conversation. A good programming Prompt needs to include information across multiple dimensions: technology stack constraints (e.g., using Spring Boot 3.x + MyBatis-Plus), coding standards (e.g., naming conventions, exception handling approaches), business context (e.g., the functional positioning of the current module), and expected output format (e.g., whether unit tests are needed). Templatized Prompts can significantly reduce communication costs and avoid having to describe requirements from scratch every time. Common Prompt template strategies in the industry include: role setting ("You are a senior Java architect"), few-shot learning (providing example code as reference), and chain-of-thought (requiring AI to analyze before coding).

Regarding quality check automation, enterprise-level projects typically need to establish a multi-layered quality assurance system: first, static code analysis using tools like SonarQube and ESLint to check code standards and potential defects; second, automated testing including unit tests (JUnit/pytest), integration tests, and end-to-end tests; third, code review by senior developers who manually audit AI-generated code, focusing on business logic correctness, security vulnerabilities, and performance concerns; and finally, automated gates in CI/CD pipelines to ensure code that doesn't meet quality standards cannot be merged into the main branch. While AI-generated code has a very high syntax correctness rate, it still requires human oversight for business logic accuracy and edge case handling, which further amplifies the importance of a quality assurance system.

Full-Process Automated Development

Teacher Zhuge has a background as a senior architect at major Chinese internet companies like JD.com and Vipshop, and also runs a company specializing in enterprise LLM project implementation. His full-process automated development methodology comes from real enterprise project practice, not theoretical speculation.

The core idea of full-process automation is: deeply integrating AI into every phase of software development (requirements analysis → design → coding → testing → deployment), not just using AI in the coding phase. This end-to-end AI integration approach is what truly unleashes AI programming productivity.

Specifically, during the requirements analysis phase, AI can assist in converting Product Requirements Documents (PRDs) into technical design plans; during the design phase, AI can generate database table structures and API interface definitions based on architectural constraints; during the coding phase, AI handles the main code generation; during the testing phase, AI can automatically generate test cases and execute tests; during the deployment phase, AI can assist in writing deployment scripts and configuration files. The AI output at each phase needs to go through manual review and confirmation, forming a standardized workflow of "AI generation → manual review → confirmed execution."

Practical Advice for Developers

Don't Give Up on AI Programming Because of Failure

Many developers choose to give up after encountering setbacks in their first attempts at AI programming, returning to the old way of writing code by hand. This is actually due to a lack of proper methodology. AI programming isn't about "throwing requirements at AI and calling it done" — it requires:

• Clear context management
• Reasonable task decomposition
• Continuous Prompt optimization
• Rigorous code review

This is similar to when agile development first emerged — many teams simply eliminated documentation and planning and claimed they were "doing agile," only to end up with chaotic projects. Any new development paradigm requires supporting methodology and accumulated practice, and AI programming is no exception.

Start Small and Scale Gradually

Developers are advised to begin practicing with the following steps:

1. First, familiarize yourself with the basic usage of Claude Code on personal projects, understanding its command-line interaction mode and project indexing mechanism
2. Build your own Prompt template library, accumulating prompts for common scenarios including CRUD operations, API development, bug fixes, and other high-frequency scenarios
3. Promote standardized AI programming processes within your team, establishing unified CLAUDE.md project configuration files and code review standards
4. Gradually build a complete Harness AI engineering system, integrating AI into CI/CD pipelines and project management tools

Conclusion

AI-engineered programming isn't the future — it's happening right now. Claude Code + the Harness methodology provides developers with a clear path to implementing AI programming. The key isn't how powerful the tool itself is, but whether you've established a systematic engineering framework to harness AI's capabilities. When you master the right methodology, 80%-90% of code can indeed be delegated to AI — while you focus on the more valuable work of architecture design, business decisions, and quality control.

From a broader perspective, AI-engineered programming is redefining the talent capability model in the software development industry. Outstanding developers of the future will no longer be measured solely by coding speed and code volume, but by architectural thinking, system design ability, AI collaboration efficiency, and engineering management proficiency. Embracing this transformation early and establishing a systematic AI programming methodology will be the key for every developer to maintain competitiveness in the new era.