Expo Web Team's Month-Long Claude Code Experience: Lessons Learned
Key Points
- Claude Code excels at knowledge retrieval for unfamiliar codebases
- Context management requires active session clearing to maintain quality
- Strong codebase patterns and conventions significantly improve AI output
Summary
Expo's web team conducted an intensive month-long evaluation of Claude Code to understand its impact on development workflows. The team identified key strengths and limitations while developing practical strategies for effective AI-assisted coding.
Key Points
Where Claude Code Excels
- Knowledge retrieval for unfamiliar codebases and system context
- Well-specified tasks with clear requirements and acceptance criteria
- Codebases with strong patterns - established conventions, linting, and type checking guide better output
- Parallelizing development using Git worktrees for simultaneous branch work
- Plan mode for complex features requiring clarifying questions and context preservation
Current Limitations
- Session memory - Claude starts fresh each time, requiring repeated onboarding
- Skill application - often forgets to apply pre-packaged recipes without explicit reminders
- Context management - long sessions lead to quality deterioration and hallucinations
- Engineering judgment still required to evaluate architecture and solution quality
Practical Workarounds
- Use concise CLAUDE.md system prompts for consistent guidance
- Implement
/clearcommand after discrete tasks or export progress to markdown - Leverage MCPs (Linear, Sentry, Figma, Graphite) for enriched system context
- Manually invoke skill slash commands when specific recipes are needed