Summary of Key Points
This article focuses on the practice of "AI-native development" within enterprises, explaining why companies may feel a sense of "disconnection" when pursuing AI-native approaches (while individual use of AI can significantly improve efficiency, implementing it at an organizational level proves challenging). It then introduces SDD (Spec-Driven Development) as an organizational solution for AI-native implementation. Through practical comparisons of two tools, Spec-Kit and BMAD, the article highlights their respective features and advantages and disadvantages. Ultimately, it emphasizes that the essence of AI-native development lies in reorganizing collaboration processes—ensuring teams adapt to AI capabilities rather than simply requiring everyone to use AI tools.
Why Do Companies Feel That AI-Native Approaches Are Unappealing or Even Disruptive?
When individuals use AI to write code, they often report a doubling of efficiency. However, when companies attempt to adopt AI-native practices, they encounter a series of seemingly unappealing issues:
- How should permissions be defined? Which tasks can be automated by AI? Who will review the resulting code changes (PRs)?
- What if the AI-generated task list includes duplicates? Who will be responsible for handling failed unit tests?
- Who will be accountable if there are errors during automatic code merges? Could the AI's messaging system for requesting additional materials offend colleagues?
These seemingly trivial rules and accountability mechanisms make AI-native approaches seem less appealing, but they are essential. Individuals seek rapid results, while organizations need stable and reliable processes. Occasional breakthroughs are not enough; what matters is that AI operates without errors and within a controlled framework. Therefore, these "unappealing" details are actually critical for the successful implementation of AI-native strategies.
The Core of AI-Native Development: Not "Everyone Using AI," but "AI Participating in Collaboration"
Many companies mistakenly believe that having everyone use tools like ChatGPT to write code constitutes AI-native development. However, the true essence of AI-native development is about integrating AI into the team's workflow, collaboration processes, and delivery mechanisms, effectively making it a virtual member of the team. For example, an AI in a practical case could:
- Automatically monitor group chats and logs to identify issues and generate task lists;
- Write code, perform tests, and submit code changes (PRs) on its own;
- Proactively seek information from responsible parties when context is lacking;
- Even use multiple models to jointly review low-risk code changes.
The critical shift is from asking "Can AI help me write the code?" to "According to what rules and business logic should AI write the code?" It is essential to provide AI with clear guidelines, requirements, and a framework to ensure consistent and reliable outcomes.
SDD: Spec-Driven Development—Setting Rules for AI
SDD is a method of defining rules for AI development. By using clear specifications (requirements, solutions, tasks), it ensures that AI operates within a structured context, reducing rework and internal inefficiencies. GitHub's Spec-Kit has popularized this approach:
- Core Function: Providing a stable environment with defined requirements, code standards, and business constraints, making organizational use of AI more orderly.
- Essence: Transforming AI from a "freely expressive assistant" into an employee that follows established rules, providing a unified basis for team collaboration.
Spec-Kit vs. BMAD: A Practical Comparison of Two Tools
1. Spec-Kit: Like a "Fixed Process Template"
- Features: Clear process (define → plan → tasks → implement) with mandatory standardization.
- Issues: In complex projects with multiple repositories and roles, additional mechanisms are required (e.g., adapting to different repository structures). For example, users must encapsulate commands to inform AI about the location of front-end and back-end code and check for API conflicts.
2. BMAD: Like an "AI Virtual Team"
- Features: Incorporates AI roles such as product managers, architects, and QA personnel, with a collaborative review process (multiple stakeholders providing feedback).
- Issues: The review workload is substantial, and it can be mentally demanding, especially for those unfamiliar with certain areas (e.g., reviewing back-end solutions by front-end developers). The initial setup phase may even take longer than traditional human collaboration.
Choosing the Right Tool Depends on the Team
- BMAD is suitable for: Single individuals or small teams lacking specific roles (e.g., without an architect), where AI can help fill in capability gaps.
- Spec-Kit is suitable for: Teams with strong capabilities and well-established processes (with clear guidelines and multiple repositories) that require more controllable AI integration.
In conclusion, AI-native development is not about accumulating various tools but about reorganizing work processes. Only by adapting teams to the capabilities of AI can we truly transform its efficiency into organizational benefits.