Code with Codex
How I use Codex as a coding partner for planning, implementation, verification, and polishing a personal portfolio.
Working with Codex feels different from using a normal autocomplete tool. It is closer to having a coding partner that can read the project, make changes, run checks, and keep the work moving while I stay focused on direction and judgment.
The most useful part is not just speed. It is the feedback loop: describe the change, inspect the result, adjust the details, verify it, and repeat.
My Workflow
I usually start with a rough intention instead of a perfect specification. For example, I might say that a blog section feels repetitive, a button is not obvious enough, or a homepage needs to feel more professional.
Codex then helps translate that into concrete implementation steps: inspect the code, identify the right files, edit the UI, run lint/build, and preview the page locally.
Where It Helps
Codex is strong at repetitive engineering work. It can update copy, add routes, wire MDX posts into a registry, adjust Tailwind classes, and run verification commands without losing the broader thread.
It is also useful for turning project code into writing. Instead of only listing technologies, it can help extract architecture, tradeoffs, and knowledge points from the repo.
Where I Still Need to Lead
The human part still matters. I need to decide what is true, what should be shown publicly, which projects are finished enough to feature, and whether the tone feels like me.
Design judgment also stays human. Codex can propose layouts, but I still need to decide when a page feels too crowded, too repetitive, or too generic.
Knowledge Points
- AI agents work best when they can inspect the real codebase instead of guessing from a vague description.
- Fast iteration is only useful when paired with verification such as lint, build, and local preview checks.
- Good collaboration means giving direction, reviewing output, and correcting the agent when the product story is wrong.
- AI can accelerate implementation, but ownership of taste, accuracy, and authenticity stays with the developer.
- A portfolio built with an AI agent should still feel personal, not automatically generated.
Takeaway
Codex helps me move faster, but the best results come from collaboration. The agent handles a lot of implementation detail, while I keep responsibility for the decisions that make the work honest and useful.
和 Codex 一起工作，感觉和用普通的代码补全工具完全不同。更像是有了一个能读懂项目、直接修改代码、跑验证脚本，同时让我保持对方向和判断的掌控的编程伙伴。
最有价值的不只是速度，而是整个反馈循环：描述意图、检查结果、调整细节、验证效果，再继续。
我的工作流
我通常从一个粗略的想法开始，而不是写好完整的需求说明。比如，我可能说某个博客区块看起来太重复，某个按钮不够明显，或者首页整体感觉还不够专业。
Codex 会把这些转化成具体的实现步骤：读取代码、定位文件、修改 UI、运行 lint/build、在本地预览效果。
它的优势
Codex 在重复性工程工作上非常擅长。它可以更新文案、添加路由、把 MDX 文章注册进博客系统、调整 Tailwind 类，并执行验证命令，而不会丢失更大的上下文。
它也很适合把项目代码转化成文字。不只是列举用了什么技术，而是从仓库里提取架构决策、取舍权衡和可迁移的经验。
哪些地方还需要我来主导
人的判断不可替代。我需要决定什么是真实的，哪些内容适合公开展示，哪些项目足够完整值得展示，整体语调是否像我自己写的。
设计判断也属于人的部分。Codex 可以提出布局方案，但我仍然需要判断页面是否太拥挤、太重复，或者太像模板。
知识点
- AI agent 能看到真实的代码库时效果最好，猜测模糊需求时容易跑偏。
- 快速迭代只有配合验证（lint、build、本地预览）才有实际价值。
- 好的协作是给出方向、审查输出、在产品叙述跑偏时及时纠正。
- AI 可以加速实现，但品味、准确性和真实感的所有权仍在开发者这里。
- 用 AI agent 构建的作品集，感觉上仍然应该是个人的，而不是自动生成的。
总结
Codex 让我迭代更快，但最好的结果来自协作。agent 处理大量实现细节，而我保留对让作品真实有价值的那些决策的责任。