When AI Reviews Code, What Should Android Engineers Review? — Building the Foundation and Redesigning Reviews
summary
An Android engineer explains how pixivコミックAndroid introduced and operated Gemini Code Assist for PR reviews, detailing how project conventions (AGENTS.md/Skills, Compose state management, DI, Kotlin style, unit/E2E tests) affected AI review accuracy and caused Android-specific failures (e.g., AI misreading generated ViewBinding/DataBinding classes). The talk also presents a redesigned review process that shifts human reviewers’ focus from coding style to specification, design, intent and risk assessment, recommends risk-based review practices (self-AI review by authors, different review intensity for UI tweaks vs. high-risk changes like billing/auth), and provides an AI-review readiness checklist and a draft for risk-based review redesign based on real operational trade-offs and failures.
AI-written summary · about
chapters (11)
- 0:00Android engineers in AI era
- 3:03Pixiv Comics Android status
- 6:57AI review failures
- 11:16Foundation with Agents MD
- 19:25Designing from failures
- 21:53Reducing load with tests
- 25:32Redesigning reviews
- 26:37Behavior vs structure changes
- 30:51Practical effects
- 32:18Learning from failures
- 34:50Review improvements
related talks
26mCode Review Around Artificial Intelligence: Good and Bad Practices
34mAI and Code Quality: Building a Synergy with Human Intelligence
19mThe AI Code Review Bottleneck: Escaping the Velocity Trap
43mCode Review in the Age of Vibe Coding
39mWhat's new in Android Studio's AI Agent?
18mAI Can End Your Career, But Not in the Way You Think
13mDeveloping with AI: Balancing Speed and Code Quality
36mCode review for Android is broken