Building a Blog with AI

Why I Built a Blog

Working in ISMS-P consulting, I'd been accumulating notes and resources that felt too valuable to leave buried in Notion. I wanted to start a blog, but with almost no development experience, I never got around to it.

Then I discovered OpenClaw, an open-source AI agent platform, and thought: "I might actually be able to pull this off."

Tools I Used

OpenClaw is an AI agent platform that provides an environment for coding through conversations with Claude. Claude (by Anthropic) handled the bulk of the code writing, debugging, and design. Docker ran the OpenClaw execution environment. GitHub managed the source code and triggered auto-deployments, Vercel handled hosting and continuous deployment, and the domain was purchased from GoDaddy.

The overall flow: OpenClaw (Docker) where Claude writes code → Push to GitHub → Vercel auto-deploys → GoDaddy domain connected.

What Went Well

Development through conversation actually works. Even without knowing how to write code, I could say things like "add a category feature to the blog" or "add dark mode," and Claude would write the code, OpenClaw would create the files, and even handle the git push. It genuinely felt like magic at first.

GitHub + Vercel auto-deployment is convenient. Push and the site updates within seconds. No server management needed, and SSL is applied automatically.

Docker eliminates environment setup. No need to install Node.js or configure a development environment on my machine. Everything runs inside the container.

It costs almost nothing. Vercel's free plan + GitHub free + just the domain cost.

AI understands the security context. When I said "create an ISMS-P category," it knew what that meant and even suggested appropriate subcategories.

What Was Difficult

Initial OpenClaw setup was rough. Docker itself was new to me, so I struggled from installation. The official documentation is written for developers, which creates a barrier for non-developers. There was some trial and error with environment variables and API key configuration.

GoDaddy DNS settings were unfamiliar. Connecting the domain to Vercel involved DNS record settings (A, CNAME) that I'd never touched before. During the propagation period, I kept wondering "did this actually work?"

AI isn't omnipotent. Sometimes Claude's code would throw errors or produce unintended results. Asking for debugging usually resolved things, but explaining what went wrong requires at least a basic development intuition.

Version control concepts can feel overwhelming. Encountering Git concepts like commit, push, and branch for the first time raised basic questions. And since AI handles it for you, there's a tendency to just delegate rather than learn.

Customization has limits. Even when I said "change it to look like this," getting the output to match my mental image 100% was difficult. Having at least basic CSS and layout knowledge would make things much smoother.

Security Considerations for OpenClaw

As someone in the security field, I feel I should address this.

OpenClaw grants the AI agent fairly broad permissions — file system access, shell command execution, external API calls, and more. It's convenient, but misconfiguration or exposure to malicious prompts could lead to unintended command execution or data leakage. Being open source, security patch cycles and vulnerability response may also be slower than commercial products.

Running inside a Docker container provides a degree of isolation, but caution is needed in environments with sensitive API keys or personal data. Best practices include using Docker for isolation, configuring API keys with least privilege, separating from networks containing sensitive data, and regularly checking for updates.

Final Verdict

AI tools made it possible to accomplish something I couldn't have done before. But "AI handles everything so you don't need to know anything" is an overstatement. You need to at least understand what Git is, what deployment means, and what DNS does in order to give AI the right instructions.

Still, it was worth doing. When you don't know something, you can just ask the AI, and the trial-and-error process itself becomes learning.

---