Teaching AI the "Manners" of Your Code ── Designing Team Development That Balances Quality and Speed

On our team, after adopting an AI coding agent, we could write code faster — but we kept running into the same problem: code that didn't follow our team's coding style triggered repeated review comments and revisions, and our reviews piled up. Even when AI speeds up the work, if your team's own style isn't conveyed to it, quality and speed remain a rade-off.

We overcame this by putting our lead engineers' tacit knowledge into words as a style guide, and teaching the AI to understand our team's coding rules. In this session, we'll share how we extracted that design philosophy, how we built AI into every step from implementation to review, and how the system keeps improving as we operate it — all backed by real data.

Rather than being tied to any specific tool, we'll introduce this as a general framework that any team can reproduce.

Transcript

1. 1 2026.06.29 LY Corporation Hideki Eguchi | Media & Search Domain, Media CTO SBU Teaching AI the "Manners" of Your Code Designing Team Development That Balances Quality and Speed AIにコードの"お作法"を教える 品質と速度を両⽴するチーム開発の設計
2. ー Joined Yahoo! JAPAN (now LY Corporation) in 2007. ー Since 2014, has been developing baseball-related services at Sportsnavi. ー Now: baseball systems × AI coding agents Hideki Eguchi
3. Agenda 1. Background & Challenge 2. Approach 3. Triple Quality Assurance 4. Results 1. 背景と課題 2. アプローチ 3. 3つの品質チェック 4. 成果 5. It Keeps Getting Better 5. 育ち続ける仕組み
4. 1. Background & Challenge Why Review Congestion Happened ｜ レビュー渋滞はなぜ起きたか
5. Bloated Frontend ̶ hard to maintain Migrating to BFF ̶ a 3-year effort, alongside daily work Adopted AI agents to speed up the migration 01 03 02 Scaling up, mid-BFF migration ｜ ⼤規模化し、BFF化の途中だった The Baseball System
6. コードは速く、レビューは遅く Faster coding, slower reviews Inconsistent manners ↓ more comments ↓ Review Congestion
7. ボトルネックが、レビューに移った Bottleneck Shifted to Review Faster ≠ better without team manners
8. 2. Approach From Tacit Knowledge to a Style Guide ｜ 暗黙知をスタイルガイドへ
9. なぜスタイルは揃わないのか Why Styles Don't Align Small habits, big gap
10. 暗黙知を、書き起こす Tacit Knowledge, Written Down Architect's Tacit Knowledge Style Guide AI Reads & Applies
11. how to code PRs opened what gets flagged コードではなく、PRを分析する Analyze the PRs, Not Just the Code PRs reviewed The reasoning lives in the PR, not the code
12. Two prompts, two angles “Analyze PRs opened by the Architect → design, rules, tests” “Analyze PRs the Architect reviewed → what they flag, the limits” Both sides reveal the full philosophy 実際のプロンプト The Actual Prompts
13. Design for Testability ̶ signs a class should be split ・ Logic you want to test is hidden in a private method ・ Mock setup is too complex → responsibilities are mixed ・ Hard to test = hard to change スタイルガイドの中⾝ Inside the Style Guide Rules from real experience
14. 3. Triple Quality Assurance Implementation, Self-Review, Review Assistance ｜ 3つの品質チェック
15. 3つの段階で品質を守る Three Stages Protect Quality Quality Guided Coding Pre-PR Review Review Assist
16. 第1段階：ガイドに沿った実装 Stage 1 : Guided Coding Quality Guided Coding Pre-PR Review Review Assist
17. 第2段階：提出前のセルフレビュー Stage 2 : Pre-PR Review Quality Guided Coding Pre-PR Review Review Assist
18. 第3段階：レビュー補助 Stage 3 : Review Assist Quality Guided Coding Pre-PR Review Review Assist
19. 4. Results Balancing Quality and Speed ｜ 品質と速度の両⽴
20. AIだけでは、⾜りなかった AI Alone Wasn't Enough
21. AIだけでは、⾜りなかった AI Alone Wasn't Enough
22. お作法が、流れを変えた The Guide Changed It
23. 5. It Keeps Getting Better 育ち続ける仕組み
24. ⾃⼰強化のループ A Self-Reinforcing Loop Review Finds an Issue AI Improves Feedback to Guide Every loop makes the guide stronger
25. Teaching AI the "Manners" ↓ Faster AND better