Introduction: The Shift in Development Focus
In the AI era, the focus of development work is undergoing a significant shift: from spending most time writing code to dedicating more effort to system architecture and design. When AI can efficiently transform designs into code, our core value is no longer writing code itself, but designing systems, defining specifications, and building knowledge systems.
This article will explore how to transform AI collaboration from inefficient "conversational flow" to efficient "documentation flow" through building knowledge base systems, truly achieving the "Architect, Don't Code" working model.
Why "Architect, Don't Code"
Current State of AI Coding Capabilities
In recent years, AI development tools have shown explosive growth. From GPT to Cursor, various intelligent programming assistants can now:
- Automatically complete most basic coding tasks
- Understand natural language requirements and convert them to code
- Handle complex code refactoring and optimization
Key Insight: Coding work occupies more than 50% of the development cycle and is also the part most easily replaced by AI.
Challenges of Traditional Development Models
However, when we still collaborate with AI as "coders," we encounter numerous problems:
1. Architecture Understanding Gaps
- AI-generated code may run but often deviates from overall architectural design
- Lacks understanding of system-wide perspective, leading to local optimization but overall chaos
- Cannot understand team design philosophy and technical selection reasoning
2. Repetitive Communication
- Need to explain architectural decisions and design principles every time
- Cannot accumulate and inherit architectural knowledge
- Significant time wasted on explaining "why design this way"
3. Quality Control Difficulties
- Inconsistent code styles
- Violations of established architectural principles
- Introduction of unnecessary technical debt
Mindset Transformation from Coder to Architect
Conversational Flow VS Documentation Flow: Work Model Innovation
Traditional Conversational Flow (Coder Mode)
Coding Requirements → Multiple Rounds of Code Discussion → Repeated Debugging and Modification → Uncontrollable Results
New Documentation Flow (Architect Mode)
Architecture Design → Specification Documentation → AI Implementation → Predictable Results
From a system design perspective, this is a transformation from "temporary instructions" to "persistent architecture." Architects define the system's "constitution" through documentation, while AI is responsible for "enforcement" — converting architecture into concrete implementation.
The Compound Effect of Design
Google's research shows that team investment in architectural design is positively correlated with final system quality. But why are most developers unwilling to invest sufficient time in design?
Psychological Explanation:
- Coding provides immediate feedback and sense of achievement
- Design value needs long-term manifestation
- Humans naturally prefer short-term gains
Opportunities in the AI Era:
When AI takes over coding work, we can finally focus our energy on architectural design that truly creates long-term value. This is not abandoning coding, but upgrading to higher-level system thinking.
Transformation of Development Processes
The Value of Design-First
The concept of "standardized, upfront design" has been validated in the software engineering field. Google's research shows:
- Team investment in design before coding is positively correlated with final code quality
- However, less than 1/3 of developers are willing to invest sufficient time in design
Cause Analysis:
- Coding brings immediate sense of achievement and feedback
- Design value manifests in later stages
- Leading to projects commonly having "insufficient early design, frequent later rework"
New Opportunities in the AI Era
The emergence of AI provides us with opportunities to change the status quo:
- Coding Work Characteristics: Occupies more than 50% of development cycles and is most easily replaced by AI
- New Work Focus: Invest more energy in early design and requirement analysis
- Expected Effects: Improve overall efficiency and make development processes smoother
Knowledge Base System
The Nature of Knowledge Bases
Knowledge bases are not a new concept, but rather systematic integration of various documents we're familiar with:
- README, CHANGELOG
- API documentation, design specifications
- Technical blogs, best practices
Core Value of Knowledge Bases
Knowledge bases act like team "collective intelligence" with three core values:
- Sustainability: Independent of specific technical frameworks or AI tools, long-term effective
- Replicability: Team experience can be systematically accumulated and reused
- Evolvability: Continuously optimized and improved through practice
Implementation Guide
Basic Setup (Building Project "Manual")
The basic setup cost for a knowledge base in a medium-scale project is approximately 1 person-day.
1. Create Knowledge Base Structure
Create a .knowledge
directory in the project root, including the following core files:
2. Write index.mdx (Basic Documentation)
## Knowledge Base Basics
- Please read and follow rules: knowledge/rule.mdx
## Project Technology Stack
- [List project dependencies and purposes, avoid AI using additional technologies]
## Project Business Background
- [Brief description of project business background]
## Project Structure Description
- utils/: Common utility functions directory
- access.ts: Permission management core logic
- [Other important directories and file descriptions]
## Business Feature List
- [Each page's path, entry file, feature description]
## Coding Standards
- [Data flow specifications]
- [Naming conventions]
- [Other team agreements]
3. Write rule.mdx (AI-Specific Rules)
## Rule Compliance
- Start each reply with "Dear Programmer, Hello"
## Code Implementation Rules
- Services divided by baseUrl, different baseUrl correspond to different files
- When developing new pages, automatically add:
- Route configuration (src/config/routes.tsx)
- Permission configuration (src/access.ts)
- Show toast notifications after successful create/update/delete operations
## Instruction Understanding
- "Complete reference": Use same components, styles, code structure and style
- When referencing images: Focus on functionality and text, styles only reference component selection
- Ignore parts marked as "Pending" in documents
Incremental Development Process
1. Create Requirement Directory
Use business-dimension directory structure:
.knowledge/
└── [business-name]/
└── features/
└── doing/
└── YYMMDD-[requirement-name]/
└── index.mdx
Example: Adding do-not-disturb list feature
.knowledge/not-disturb/features/doing/241215-add-list/index.mdx
2. Write Requirement Documentation
Part I: Background Knowledge Supplement
- Business background explanation
- Related concept interpretation
- Dependency relationship explanation
Part II: Code Implementation Plan
Break functionality into independent small requirements, handle separately:
Brand New Requirements (Adding Code)
- Prioritize using code templates to maintain style consistency
- Clearly reference pages or components
- Use coarse-grained descriptions
## Feature: User List Page
- Reference: src/pages/admin/list (identical page structure and styles)
- Location: src/pages/user/list
- Requirements: Replace corresponding APIs and fields
Iteration Requirements (Modifying Code)
- Need fine-grained descriptions
- Include all modification points
## Feature: Add Filter Conditions
- Code Location: src/pages/user/list/index.tsx
- Background: Currently only supports name search
- Implementation Plan:
1. Add status filter in SearchForm component
2. Modify API request parameters
3. Update list refresh logic
Usage Tips
Prompt Examples
Initialize Context:
[Manually add feature/index.mdx and knowledge.mdx to context]
Now please help me implement feature: User List Page
Continue Next Feature:
Okay, now continue implementing feature: Add Filter Functionality
Notes
- Verify Rule Execution: Check if AI replies include agreed identifiers (like "Dear Programmer, Hello")
- Handle Context Loss: Manually reintroduce relevant documents when necessary
- Handle Call Limits: When exceeding 25 calls, input "Please continue"
- Configure Auto-Import: Configure rules in Cursor to auto-import knowledge base
Future Outlook
Public Knowledge Base Construction
- Cross-project public instruction library
- Best practice template library
- Unified knowledge indexing system
Intelligence Enhancement
- Automated knowledge extraction
- Intelligent document recommendations
- Dynamic update optimization strategies
We have established a sustainable, replicable, and evolvable knowledge system. This will become a valuable asset for long-term team development, helping us maintain competitive advantages in the AI era.
Top comments (6)
Love the focus on shifting from code to architecture and designing for AI - did you find it tough to get your team invested in documentation at first, or did the benefits make it click quickly?
Once someone showed that it only took 2 hours to do a day's work, there was a steady stream of people interested and trying it.
that's a fantastic article, i do not compliment.
bro thank you so much
Looking forward to your reply after practice
Love how this breaks it all down to actual steps you can follow - always felt like having docs saves me later.
And now we don't need to waste time on code, which takes more than 50% of our time