[mercari GEARS 2025] Specs to Code with Coding Agents: Where Do Engineers Come In?

Transcript

1. Specs to Code with Coding Agents: Where Do Engineers Come In?
 Toshiki Kawamura
 Mercoin / Backend Engineer

2. Toshiki Kawamura
 
 
 I started my career as a software engineer at a major real estate tech company in 2019 and later joined a startup, gaining broad experience in backend development. I joined Mercari in 2022, where I have led the development and operation of crypto currency trading services.
 Mercoin / Backend Engineer

3. When AI Codes, What’s Left for Us? • As AI coding becomes common, I wondered if clear specs alone could let AI build the code.
 • So I tried spec-driven development on a real project.
 • In this talk, I’ll share what worked, what didn’t, and what I learned about how engineers and AI can work together.
 

4. What I Actually Tried
 Coding Agent (Claude Code)
 • Used MCP + Coding Agent (Claude Code)
 • Asked clarifying questions and updated specs
 • Generated initial code → ran tests → refactored

5. Result
 
 Fast, But Not “Good”
 • ✅ AI quickly generated code that worked according to specs
 • ❌ However, generated code wasn’t necessarily ‘good’ code
 ◦ Good code = Readable, maintainable, well-designed
 • Needed heavy refactoring before use
 📊 76% of developers refactor at least half of AI-generated code before using it. Source: State of AI 2025 Survey https://2025.stateofai.dev/en-US/usage/#ai_generated_code_refact
 

6. Why Can't AI Write "Good Code" with Spec ?
 
 
 AI lacks critical context
 • Domain rules and business context
 • The “why” behind each feature
 • Organizational tacit knowledge
 • Technical trade-offs
 → I needed to provide this missing knowledge and context myself

7. Attempting Improvement
 
 • Followed a structured workflow inspired by GitHub Spec Kit: specify → plan → task →implement
 • Switched from abstract to step-by-step prompts with examples
 • Asked AI to explain its reasoning before coding

8. Insights from Improvement
 
 
 
 What I Learned
 • To guide AI, engineers must first design clearly in their minds.
 • Explaining specs to AI requires deep product and domain understanding.
 • AI makes us faster, but doesn't inherently make us better in the long term.

9. Where Do Engineers Come In?
 
 
 
 AI lowers coding cost → more code → more complexity.
 • As AI evolves, two abilities become more important than ever:
 ◦ System design
 ◦ Business and domain knowledge — communicating context to guide AI

10. Closing
 
 
 
 • AI is accelerating development, but our role is more essential than ever.
 • The more AI advances, the higher the level of judgment we need.
 • There’s still so much to build — and even more to learn.
 　　　 🐎 Let’s harness AI and move things forward.
 
 

11. Thank You!