Over the past year, the phrase “I wrote it with AI” has gone from “worth mentioning” to “weird if you didn’t” in pull request descriptions. Anthropic internally merged a single PR of 22,000 lines — largely produced by Claude, landing in their production reinforcement-learning codebase. When that story spread, most engineers didn’t react with awe; they reacted with anxiety: if they can pull that off, what’s stopping the wave of “PMs opening Claude Code and writing production code” from hitting our team?

You can’t stop it. But you can manage it. This post is the SDLC guideline I wrote for my own team after working through Erik Schluntz’s talk Vibe coding in prod | Code w/ Claude (Erik is a coding-agent researcher at Anthropic and co-author of Building Effective Agents). The whole thesis is one sentence: enjoy the AI speedup without sacrificing code quality, maintainability, system reliability, or security.

過去一年，「我用 AI 寫的」這句話在 Pull Request 描述裡出現的頻率，已經從「值得一提」變成「沒寫才奇怪」。Anthropic 內部曾合併過一個 22,000 行、由 Claude 大量產出、最後跑進正式環境強化學習程式碼庫的 PR——這件事傳出來的時候，多數工程師的第一反應不是讚嘆，而是焦慮：如果他們做得到，那我們團隊憑什麼擋得住「PM 直接開 Claude Code 寫 production code」的浪潮？

擋不住。但可以管好。這篇文章是我整理 Erik Schluntz（Anthropic 編程智能體研究員、《Building Effective Agents》共同作者）演講「Vibe coding in prod | Code w/ Claude」之後，為自己團隊寫的一份 SDLC 規範。核心命題只有一句：在享受 AI 加速的同時，不犧牲程式品質、可維護性、系統穩定性與資安。

Most teams meet Kubernetes for the first time through a single “superuser” kubeconfig — a file with credentials that can wipe out the entire cluster. It then starts bouncing around Slack, email, and laptops, and nobody can say for sure who still has a copy, or whether the kubeconfig belonging to a former employee is still usable.

This post shows how to use kubelogin together with AuthGate to stand up an OIDC login flow on k3s: the moment a user runs kubectl get pods, the browser pops open AuthGate’s login page, the tokens land back in the kubeconfig, and the entire cluster no longer needs that shared admin.kubeconfig.

大部分團隊第一次裝 Kubernetes 時，拿到的都是一個「超級使用者」的 kubeconfig，裡面就躺著一組可以幹掉整個叢集的憑證。於是這份檔案開始在 Slack、Email、筆電之間被複製來複製去，沒有人知道目前誰還留著副本、哪個離職員工的 kubeconfig 還能用。

這篇文章要示範如何用 kubelogin 搭配 AuthGate，在 k3s 上建立一條 OIDC 登入流程：使用者打 kubectl get pods 的瞬間，瀏覽器自動跳出 AuthGate 的登入頁面，登入完成 token 寫回 kubeconfig，整個叢集不再需要共用那份 admin.kubeconfig。

Gitea Act Runner is the execution component of Gitea Actions, responsible for fetching CI/CD tasks from the Gitea Server and reporting execution results. As more teams self-host Gitea, the HTTP request volume between Runners and the Server has become a bottleneck on the Server side. This article documents how we analyzed and resolved this problem, reducing the request volume from approximately 1,300 req/s to approximately 170 req/s for 200 Runners — an 87% reduction.

Update (2026-04-20): This article originally described the design after PR #819 merged. A follow-up, PR #822, revealed during code review that #819 had introduced a concurrency regression for Runners with capacity > 1, and addressed it with a “single poller + semaphore” architecture. See the new section at the end: Follow-up: Single Poller with Semaphore (PR #822).

Gitea Act Runner 是 Gitea Actions 的執行元件，負責從 Gitea Server 領取 CI/CD 任務並回報執行結果。隨著越來越多團隊自架 Gitea，Runner 與 Server 之間的 HTTP 請求量成為了 Server 端的瓶頸。本文記錄我們如何分析並解決這個問題，將 200 個 Runner 的請求量從約 1,300 req/s 降到約 170 req/s，降幅 87%。

更新（2026-04-20）：本文原本描述 PR #819 合併後的設計。後續 PR #822 在 code review 中揭露 #819 對 capacity > 1 的 Runner 引入了一個 concurrency regression，並用「單一 poller + semaphore」架構再次修正。請見文末〈後續修正：Single Poller with Semaphore（PR #822）〉章節。

In previous posts, I introduced the concept of Agent Skills and showed how to build an AI-driven development workflow with Claude Code + GitHub Copilot Review. As more products and teams embrace AI Agents, a clear architectural pattern is emerging: API + CLI + Skills. This isn’t a framework or protocol — it’s a pragmatic three-layer architecture that enables any product to become “agent-friendly” quickly.

在之前的文章中，我分別介紹了 Agent Skill 的概念以及如何用 Claude Code + GitHub Copilot Review 打造 AI 驅動的開發流程。隨著越來越多產品和團隊開始擁抱 AI Agent，一個清晰的架構模式正在浮現：API + CLI + Skills。這不是某個框架或協議，而是一種務實的三層架構，讓任何產品都能快速變得「對 Agent 友善」。

As a developer who maintains multiple open-source projects on GitHub, my daily work goes far beyond writing code. I also deal with a constant stream of issues, PR reviews, releases, and other tasks that are tedious but critical. As projects grow, this workload quickly exceeds what one person can handle efficiently.

Over the past few months, I’ve integrated Claude Code and GitHub Copilot Review into my daily development workflow, and the results have been remarkable — tasks that used to take half a day now often get done in 1-2 hours. This post shares my complete workflow and explains why I believe the developer’s own technical ability is the key to using AI tools effectively.

身為一個在 GitHub 上長期維護多個開源專案的開發者，我每天面對的不只是寫程式，還有大量的 Issue 處理、PR Review、版本發佈等瑣碎但重要的工作。隨著專案規模成長，這些工作量已經遠超一個人能高效處理的範圍。

過去一段時間，我開始將 Claude Code 和 GitHub Copilot Review 整合進我的日常開發流程，結果讓我非常驚艷——原本需要半天的工作，現在經常在 1-2 小時內就能完成。這篇文章將分享我的完整流程，以及為什麼我認為開發者本身的技術能力，才是善用 AI 工具的關鍵。