Introduction

In this article, I’ll share my experience as a Ruby on Rails developer using various AI coding tools (Cursor, Devin, and Cline) in real-world projects, evaluating their strengths, weaknesses, and impact on practical productivity.

Development Environment and Previous Tool Experience

I’ve been working with AI coding tools in a Ruby on Rails 8 codebase. Until recently, I primarily used Cursor with either Gemini 2.5 or Claude Sonnet 3.7 models.

Cursor Usage Patterns

  • Breaking down coding tasks into smaller requests rather than generating large code blocks at once
  • Adopting a step-by-step approach for stable “vibe coding”
  • Real-time interaction for code generation and modification
  • Using code autocomplete for quickly implementing routine code

Devin Experience

Why I Got Interested in Devin

Recently, Devin has gained attention as an AI coding tool that positions itself as a ‘software engineer capable of independent thought’. What particularly attracted me was its ability to work asynchronously on tasks with its own judgment after receiving instructions.

Pricing Structure

  • The entry barrier has been lowered from $500/month to a minimum of $20 (9 ACUs)
  • Charging is based on consumption of ACUs (Devin’s work tokens)
  • Additional purchases start from a minimum of 10 ACUs ($22.5)

Test Tasks and ACU Consumption

I performed four test tasks in a personal project:

  1. Random Code Refactoring: 0.99 ACUs
  2. GitHub Action Workflow Creation: 0.55 ACUs
  3. Complex Controller Method Refactoring: 3.14 ACUs
  4. GitHub Action CI Workflow Implementation: 2.18 ACUs

Performance Evaluation

  • At $2.25 per ACU, the cost is considerably high
  • The refactoring results did not meet expectations, especially considering the cost.
  • Failed to resolve MySQL connection issues (mysql_native_password setting) when using the trilogy adapter in Ruby on Rails
    • Devin couldn’t solve a problem that a human could fix with a single line of code

Reconsidering Cline

Why I Returned to Cline

After using Devin, I reconsidered Cline as an alternative. Previously, I judged it unsuitable as a primary tool due to its lack of code autocompletion features, but I decided to reevaluate its potential as an alternative to Devin.

New Usage Approach

  • Delegating secondary, minor tasks to Cline while focusing on main development myself
  • Interacting in Plan mode and then reviewing results from Act mode as PRs
  • Building an efficient collaboration model with partial human involvement

Test Tasks

  1. CI Implementation (where Devin failed): Successfully completed with human intervention
  2. Specific Service Class Improvement and Test Writing: Achieved satisfactory results

Conclusion and Optimal Usage Strategy

Devin Evaluation

Devin still has limitations in performing a “full developer’s role.” While there is some productivity improvement when humans review and enhance initial results, the need to spend time crafting effective prompts means its core workflow isn’t significantly different from Cursor’s. Considering its cost-efficiency, the value of its continued use appears limited.

Cline + Cursor Combination Strategy

  • Cursor: Primary development work and real-time code support
  • Cline: Parallel processing of secondary tasks and PR generation
  • Task Separation: Separating work areas by independent folders/files to prevent conflicts

This combination allows developers to focus on their main tasks while AI tools handle secondary tasks in parallel, creating an efficient workflow.

AI Coding Tools Comparison

Here’s a comparison of AI coding tools based on my usage patterns:

Feature Cursor Devin Cline
Overview AI assistant integrated into code editor Autonomous software development agent AI teammate collaborating with developers
Operation Mode Real-time interaction based Asynchronous autonomous work Hybrid approach with Plan/Act modes
Base Model Options for Gemini 2.5 or Claude Sonnet 3.7 Proprietary custom model Options including Claude Sonnet 3.7
Key Advantages • Immediate code autocompletion
• Easy use through editor integration
• Efficient for small-scale tasks
• Real-time feedback		 • Asynchronous task execution
• Autonomous task handling
• Ability to comprehend the codebase
• GitHub workflow automation		 • Detailed planning (Plan mode)
• Automatic PR generation
• Balance between interaction and autonomy
• Collaborative experience like a team member
Key Disadvantages • Limited large-scale code generation
• No asynchronous work capability
• Difficulty handling multiple tasks		 • High cost
• Complex environment setup issues
• Limited ability for detailed problem-solving
• Requires human verification		 • Lack of code autocompletion features
• Limitations with some complex tasks
• Requires separate editor usage
Pricing Structure $20/month (personal) Minimum $20 (9 ACUs) with
1 ACU = $2.25		 Usage-based (requires own Claude API Key)
Optimal Use Cases • Small to medium code writing
• Bug fixing
• Code refactoring
• Developer-led tasks		 • Simple feature implementation
• Test automation
• Independent task execution		 • Simple feature implementation
• Test automation
• Plan-based development work

CodeRabbit: Additional Verification for AI Coding Tools

The code review tool CodeRabbit is excellent not only for individual developers to review their own work but also for verifying code written by Devin or Cline. While working on the tasks mentioned above, I compared the results with CodeRabbit reviews connected to the repository. It pointed out areas needing improvement while also acknowledging well-written sections, which increased my confidence in the AI-generated code.

Additional Considerations and Future Outlook

While AI coding tools are rapidly evolving, they currently work best as collaboration tools that enhance human developers’ productivity rather than completely replacing human developers. Human judgment and intervention remain essential, especially for complex problem-solving and high-level system design tasks.

In the future, as these tools improve in cost-efficiency and domain-specific knowledge, they could contribute even more to development productivity. For now, understanding each tool’s strengths and combining them appropriately seems to be the most effective strategy.