Field Notes #4
A practical use of AI
It seems AI is all anyone can talk about on LinkedIn these days. So many strong feelings and dramatic claims. Meanwhile many of us are using AI in very practical ways. One way you'll never see directly is in helping me manage Aevocam.com's public website.
I've never been a big fan of off the shelf content management systems (CMSes). I'm a programmer. I build my own and integrate it to whatever project it's part of. I usually keep it simple. In the case of Aevocam.com, all the pages are formatted as standard Markdown text with some custom tags I added. The pages are simple, easy to read, and very SEO friendly.
I want to spend more time writing code than maintaining the website. So I created a custom agent to help me with content management. I'm using the "GPT-5 Mini and "GPT-5" models under the hood. The agent is there when I'm searching for articles (pages) and when I'm editing one. It can do the same basic things a coding agent can do: search for articles and read and write them. Its changes are in draft form for me to review and then publish.
I also added some extra tools to help the agent. It can request a list of broken links in pages or ask for a hierarchic sitemap built from links in the articles, for example. But the LLM is doing the real work. And I'm orchestrating the process.
It didn't take long to build this agent. A few days to build most of it and occasional sessions to add features. But it's totally that small time investment. It has helped me rework the entire structure and rewrite many of the pages to be a more coherent part of the larger site. It helps me keep external links fresh by finding good replacements as needed.
My little agent does incur token usage fees. I made sure to capture usage data to save in my database of current and past conversations. I'd say my typical productive conversations for fixing an article or drafting something new costs me 1 - 10 cents each. I don't think any conversation has cost me anywhere near $1 yet. This is a bargain compared to paying a human to do the same work I ask of my agent.
One huge thing worth noting is that I still turn to my paid ChatGPT account for some of my content management work. The main reason is that ChatGPT really does learn more about you over time. Your past conversations become input for later conversations. Which is amazing, really. When you integrate a GPT model like with my agent, you don't get that capability with it. Each conversation lives in a vacuum from the others. I opted not to try to build any sort of persistent "memory" for now. Trade-offs. But it means I end up using both tools.
Overall this has been totally worth the small effort needed to build it. It's saved me a lot of time and helped me with a necessary chore I don't want to spend much time doing.