May 2026 · 10 min read
Here's a question people keep asking me: "What AI tools do you actually use?"
The answer is simpler than you'd think. I don't have some secret stack of fancy AI tools. I use three things — and that's it. The magic isn't in the tools. It's in how you chain them together.
This is my primary content engine. I use DeepSeek for almost everything related to writing: articles, meta descriptions, image prompts, even the funny copy on the roast pages.
Why DeepSeek and not ChatGPT? Two reasons. First, it's significantly cheaper — about 1/10th the cost of GPT-4. Second, the Chinese-language output is noticeably better (makes sense, since it's built by a Chinese company). For bilingual sites like the food guide and travel guide, this matters a lot.
Real prompt example (for the food guide's Chengdu page):
"Write a bilingual (English + Chinese) food guide page for Chengdu, Sichuan. Target: Western travelers visiting China. Include: (1) 3 must-try local dishes with ordering tips, (2) cultural context for each dish, (3) honest truth about what actually tastes great vs what's tourist hype. Tone: warm, personal, honest. English 600-800 words, Chinese 400-600 characters."
That single prompt gives me about 80% of a finished page. I then edit and localize.
I can't code. At all. Every line of HTML, CSS, and JavaScript across all my sites was written by an AI coding assistant. I just describe what I want, it writes the code, I test it, and we iterate.
Real workflow:
This is not "prompt engineering." This is just telling a tool what you need, like you'd tell a human developer. The difference is, the AI doesn't get annoyed when you change your mind 5 times.
Hosting is free. GitHub Pages costs $0. Cloudflare's free tier handles DNS and CDN. Domain names cost about $1-2 each on NameSilo.
My total infrastructure cost so far: about $10. That's 5 domain names. Everything else is free.
Deployment is dead simple. I write code locally, push to GitHub, and GitHub Pages automatically deploys. For the AI site, I use Cloudflare Pages directly via their CLI tool (Wrangler) — one command and it's live.
Here's what nobody tells you about using AI to build websites: the tools are not the bottleneck. You are.
The AI can write the code. It can write the content. It can even suggest design improvements. But someone still needs to:
That someone is me. And that's the part AI can't replace. Yet.
If I were starting over, I'd focus on one site at a time instead of building 5 simultaneously. Spreading myself thin meant none of them got enough content in the first month. Also, I'd spend more time on keyword research upfront — I picked topics I liked (Sichuan food, city travel) rather than topics with proven search demand.
Lesson learned: passion matters, but data matters more.
— Kuang Shan, documenting the process as I go
2026年5月 · 阅读约10分钟
有个问题一直有人问我："你实际用了哪些AI工具？"
答案比你想象的简单。我没有一套花哨的秘密工具。我就用三个——仅此而已。魔力不在工具本身，在于你怎么把它们串起来。
这是我的内容主力引擎。几乎所有跟写作相关的事都用它：文章、元描述、图片提示词、甚至吐槽页面上那些搞笑的文案。
为什么不用ChatGPT？两个原因。第一，便宜得多——大概是GPT-4的十分之一。第二，中文输出质量明显更好（毕竟是国产的）。对于美食指南、旅游指南这类双语站，这个优势很重要。
实际使用的提示词示例（美食站成都页）：
"写一篇成都美食的双语指南（英文+中文）。目标读者：来中国的西方游客。内容包括：(1)3道必吃本地菜+点餐技巧，(2)每道菜的文化背景，(3)实话实说——哪些真的好哪些是游客噱头。语气：温暖、个人化、诚实。英文600-800词，中文400-600字。"
单这一个提示词就能搞定页面80%的内容。我再编辑和本地化一下就行。
我不会编程。一点都不懂。我所有站上的每一行HTML、CSS和JavaScript都是AI编程助手写的。我只需要描述想要什么，它写代码，我测试，然后迭代。
实际工作流：
这不是什么"提示词工程"。就是告诉工具你需要什么，跟你跟人类开发者说的一样。区别是，AI不会因为你改了5次主意就不耐烦。
托管免费。GitHub Pages 零成本。Cloudflare 免费套餐处理DNS和CDN。域名每个大约$1-2，在NameSilo买。
到目前为止的总基础设施成本：约$10。5个域名。其他全免费。
部署极其简单。本地写代码，推GitHub，GitHub Pages自动部署。AI站我用Cloudflare Pages直接通过命令行工具（Wrangler）部署——一条命令，立马上线。
没人告诉你的是：工具不是瓶颈。你才是。
AI能写代码。AI能写内容。AI甚至能建议设计改进。但总得有人来：
那个人就是我。而这部分是AI还替代不了的。至少目前还不行。
如果可以重来，我会一次只做一个站，而不是同时搞5个。分散精力意味着第一个月哪个站的内容都不够。另外，我会花更多时间做前期关键词研究——我选了感兴趣的话题（川菜、城市旅游），而不是有明确搜索需求的话题。
教训：热爱很重要，但数据更重要。
—— 况山，边做边记录