Draft-Final Mode

Draft first, AI refines the solution, precise code generation

Usage Share

~36% — The most used mode in our team

Mode Overview

Draft-Final Mode is our team’s most recommended approach for complex tasks. Core workflow:

1. Draft — Developer writes a rough requirement description
2. Final — AI refines it into a detailed technical solution using context retrieval, Rules, and MCP tools
3. Code — AI generates code according to the Final solution

This mode gives developers global visibility before code generation, allowing solution adjustments before implementation.

Flow Diagram

Use Cases

Why It’s the Most Recommended Mode

Comparison with Document Mode

Value of Draft

Efficiency Benefits

In our practice, using Draft-Final Mode:

• Average 3 requests to complete a complex task
• Reduced rework — Finding issues in Final phase costs less than in code phase
• Improved consistency — Final solution ensures implementation matches expectations

Complete Example

Step 1: Write draft.md

# Order Export Feature
 
## Requirement
Users want to export order list as Excel file
 
## Initial Thoughts
- Support time range filtering
- Support selecting export fields
- Need async processing for large data
- Email notification when export completes
 
## Questions
- Unsure which library to use for Excel generation
- Unsure how to handle async tasks
- Need to consider export permissions

Step 2: Generate final.md

AI will automatically retrieve project context and generate a final.md like this:

# Order Export Feature - Technical Solution
 
## Technology Choices
- Excel generation: Use exceljs (existing project dependency)
- Async tasks: Use existing Bull queue
- File storage: Reuse existing S3 configuration
 
## Data Model
 
### ExportTask Table
​```sql
CREATE TABLE export_tasks (
  id SERIAL PRIMARY KEY,
  user_id INTEGER REFERENCES users(id),
  status VARCHAR(20) DEFAULT 'pending',
  file_url TEXT,
  created_at TIMESTAMP DEFAULT NOW()
);
​```
 
## API Design
 
### POST /api/orders/export
Create export task
 
**Request:**
​```json
{
  "startDate": "2024-01-01",
  "endDate": "2024-01-31",
  "fields": ["orderId", "customerName", "amount"]
}
​```
 
**Response:**
​```json
{
  "taskId": "export_123",
  "status": "pending"
}
​```
 
### GET /api/orders/export/:taskId
Query export task status
 
## Implementation Steps
 
1. Create ExportTask data model
2. Implement export task creation endpoint
3. Implement Bull queue consumer
4. Implement Excel generation logic
5. Implement file upload to S3
6. Implement email notification
7. Implement frontend export button and progress display
 
## Permission Control
- Reuse existing @middleware/auth.ts authentication middleware
- Add 'order:export' permission check
 
## Notes
- Maximum 10000 records per export
- Files auto-deleted after 7 days
- Failed tasks auto-retry 3 times

Step 3: Review and Adjust Final

Before generating code, developers can:

• Review if technology choices are appropriate
• Adjust API design
• Add missing edge cases
• Modify implementation step order

Step 4: Generate Code

Best Practices

Draft Writing Tips

✅ Good Draft
# Feature Name
## Requirement Background (why)
## Core Features (what)
## Initial Thoughts (how - rough ideas)
## Uncertain Points (need AI to supplement)
 
❌ Bad Draft
- Too simple: Only one sentence
- Too detailed: Written as complete solution (then just use Document Mode)

Final Review Checklist

• Technology choices align with project standards
• API design follows RESTful conventions
• Data model is reasonable
• Implementation steps are executable
• Edge cases are fully considered
• Permissions and security are handled

File Organization

specs/
├── 001-order-export/
│   ├── draft.md          # Original requirement draft
│   ├── final.md          # AI-refined technical solution
│   └── api-spec.json     # Optional: detailed API spec
├── 002-user-auth/
│   ├── draft.md
│   └── final.md
└── ...

FAQ

Q: What if the generated Final solution isn’t satisfactory?

Continue the conversation to adjust:

Q: When to use Document Mode instead of Draft-Final?

When you already have a clear, complete technical solution, using Document Mode directly is more efficient. The value of Draft-Final Mode is letting AI help refine the solution.

Q: How detailed should the Draft be?

Draft should include:

• Core goals of the requirement
• Your initial approach
• Uncertain points

No need to include complete technical details — that’s Final’s job.

Next Steps

Our Draft-Final workflow embodies the principles of Spec-Driven Development (SDD) without relying on external frameworks. Learn more about Why No Spec Framework and how we’ve adapted these principles to our enterprise context.

Then, let’s see how to integrate all modes into a complete Workflow Practice.