What We Should Learn from Claude Code

ROSCAFE: https://rosca.connpass.com/event/363169/

Slide in Japanese: https://speakerdeck.com/s4yuba/claude-codekarawo-gaxue-bubekikoto

X(Twitter)：https://x.com/gaishi_narou

Transcript

1. What We Should Learn from Claude Code (2025/08/05) #ROSCAFE Speaker: Oikon
2. Self Introduction Oikon Software Engineer Foreign IT Company R&D Hobbies: Personal Development, AI Tool Research 𝕏 (@gaishi_narou): Zenn:
3. Today’s Topic ❌ Claude Code Tips ⭕️ What We Can Learn from Claude Code
4. Claude Code Tips Are Everywhere Claude Code を初めて使う人向けの実践ガイド | morizyun氏 人に寄り添うAIエージェントとアーキテクチャ | 名村卓氏 Claude Codeを実務開発で使い倒して得られた知見 | 鹿野壮氏 Claude Code: Best Practices and Pro Tips | HTDOCS.DEV テストから始めるAgentic Coding 〜Claude Codeと共に行うTDD〜 / Agentic Coding starts with testing | r.kagaya氏 初学者でも今すぐできる、Claude Codeの生産性を10倍上げるTips | Oikon and more…
5. Claude Code’s changes CHANGELOG.md Update History Major version 1.0.0 launched on May 23 Already at 1.0.68 as of August 5 (updated this morning) Risks of Over-customizing CLAUDE.md and Subagents Behavior changes over time, so over-customization may become unusable "Claude Code Has Become Dumb" Rumors Context compression specification changes are causing issues Anthropic could change service content anytime. Detailed tips learned today may become obsolete quickly.
6. Claude Code Version Changes
7. What We Can Learn from Claude Code? Focus on AI models and coding tools as a whole, not just tips ① Current Context Engineering Challenges ② Mindset for Working with AI Agents ③ Anthropic’s Vision for Claude Code’s Role
8. ① Current Context Engineering Challenges Importance of Maintaining Clean Context "Noise removal" is essential to prevent AI confusion. Hints: Context compression of proper nouns (kent-beck, t_wada), Serena LSP integration, etc. Limited Context Window Size Claude Code: 200k tokens, Gemini 2.5 Pro: 1M tokens Large differences exist between AI models, so give instructions consciously Rather than blindly requesting tasks, understand how current AI models handle context [1] [[1] Serena](https://github.com/oraios/serena)
9. ② Mindset for Working with AI Agents Praying to Claude Code Claude Code may produce better results through multiple attempts AI agents have uncertainty, so don’t fixate on completing tasks in one shot Understand That Engineering Work Is Shifting Upstream Engineers are shifting from coding tasks to designing "what" should be built To work with AI agents, engineers ultimately need to improve their fundamental skills
10. ③ Anthropic’s Vision for Claude Code’s Role "Claude Code is like that coworker that does everything in the terminal and never touches a GUI" Anthropic aims for AI agents that can be fully entrusted with tasks While not possible to fully delegate work currently, it will become possible with AI model evolution in the near future As autonomous AI agents evolve, we need to keep in mind what kind of work we can do [1] [[1] Claude Code Best Practices](https://www.youtube.com/watch?v=gv0WHhKelSE)
11. A Slide from Claude Code Best Practices (5/22) [[1] Claude Code Best Practices](https://www.youtube.com/watch?v=gv0WHhKelSE)
12. Summary Claude Code’s specifications change rapidly; tips won’t remain useful forever Learning the AI development trends from Claude Code as a tool is impoortant ① Current context engineering challenges ② Mindset for working with AI agents ③ Anthropic’s vision for Claude Code’s role We need to understand the essence of AI evolution while becoming great engineers I'll share the slides, please follow me! @gaishi_narou
13. References Claude Code を初めて使う人向けの実践ガイド | morizyun氏 人に寄り添うAIエージェントとアーキテクチャ | 名村卓氏 Claude Codeを実務開発で使い倒して得られた知見 | 鹿野壮氏 Claude Code: Best Practices and Pro Tips | HTDOCS.DEV テストから始めるAgentic Coding 〜Claude Codeと共に行うTDD〜 / Agentic Coding starts with testing | r.kagaya氏 初学者でも今すぐできる、Claude Codeの生産性を10倍上げるTips | Oikon oraios/serena: Language Server Protocol for AI Agentic Coding The new era of software development | Anthropic