Tag
观察到许多独立开发产品的开发者已停止直接使用 git，而是让 Claude 或 Codex 代办所有版本控制工作，传统的开发"肌肉记忆"正在迅速消退。
Imprint shares how its `small-change` approach speeds up pull request review, explaining eligibility criteria and the reasoning behind rethinking review purposes as PR volume and review times grow.
OpenAI has published examples of using Codex for app maintenance. The workflow uses AI to detect candidates for unused code, verify them against actual runtime data, and generate small, incremental fix proposals — with humans handling the final review.
Matt Pocock 分享了一个 prompt，用于创建 pre-commit 钩子：限制文件超过 5,000 tokens（按 bytes/4 计算）时提交失败，仅检查暂存文件并忽略非 AI 生成的文件，同时在错误信息中提供拆分文件的明确指引。
一篇关于「脑三明治」工作流的博客文章，提出工程师在使用 AI 编码时应先用脑理解问题、再放手让 AI 执行、最后再用脑审查成果，以避免技能退化并保持代码可审查性。
The author shares a human-in-the-loop continuous development workflow based on Claude Code as the main agent orchestrating Codex CLI, Subagents, and GitHub Issues/PRs, covering repo initialization conventions, task decomposition, and review/merge processes, along with reusable prompt templates.
一名工程师分享个人感悟：意识到自己并不比 AI 更聪明，未来应专注于成为 AI 的优秀编排者，为其提供正确的工具、数据与权限，同时承认自己比 AI 更懂的领域正在快速缩小。
A Lobsters discussion asking whether teams allow LLMs to write both production code and tests, with the author considering a policy restricting AI generation to only one of the two to avoid tautological, over-engineered results.
Simon Willison 引用了 Addy Osmani、Morgan Westlee Lunt 和 Qing Wang 关于 Anthropic Claude Code 团队代码审查实践的观点，指出 AI 审查让代码审查从发布前的关卡变为写作过程中的持续检查，从而让人类审查专注于更高层次的问题。
作者质疑在 AI 代理时代继续使用 pull request 的必要性，认为代码审查、知识共享和 CI 都可由 AI 模型与代理更高效地完成，主张转向更快速的持续集成工作流。
一位开发者发推询问大家如何管理历史 Claude Code 会话，抱怨需要手动记录 `claude resume` 命令，并指出 Codex/ChatGPT 已很好地解决了这个问题。
Matt Pocock argues that while agent skills aren't strictly necessary if you already follow a great process, they make it easier to follow that process consistently.
A viral tech take arguing that highly automated codebases exist to help humans ship product decisions faster, quoting Kent C. Dodds on shipping 60+ PRs a day solo with AI agents.
Indie game developer Kyle Kukshtel writes up his 'Your Game Is A Harness' talk from Indie Rodeo Calgary, exploring how to practically use LLMs like Claude Code in day-to-day game development rather than one-shot demos, and why LLMs struggle with game-specific code.
作者分享让持久化 AI 代理真正可用的三点经验：为每个代理分配独立容器、主动测试安全边界而非默认信任，以及定义可机器验证的"完成"标准。
Matt Pocock lists common mistakes developers make when working with AI coding agents, including misaligned effort, overly long prompts, weak automated checks, poor codebase navigation, and insufficient agent autonomy.
Kent C. Dodds argues that preventing AI 'slop' is a straightforward three-step engineering process — high-quality agent PRs, solid PR review, and feature flags with disaster recovery — noting these practices long predate AI agents.
Matt Pocock asks how developers are finding the GLOSSARY.md approach from /grill-with-docs, noting that investing more in the glossary makes him more productive and helps the AI agent 'speak his language'.
Addy Osmani emphasizes that verification is crucial for getting good results from AI coding agents: clearly define "done" before review, give the agent executable checks (bash commands, tests, browser tests) to pass, and only ship when nothing blocks merge.
A tweet sharing best practices for Claude Code's multi-agent workflow: use Sonnet to build, Haiku subagents for parallel grunt work, and Opus only at key review checkpoints to save tokens and context.