I’ve been using AI for the last two to three years, but this year I went deep and started building my own agents. This is what I learned building one of them, an agent that helps me run this website.
Starting as a non-coder
As a non-coder, using Claude Code sounded scary. I haven’t coded since my programming classes in university, but I was determined to get the full benefits of AI for myself, so I set it up in the Claude desktop app.
I started with Claude Code, and I’ll use my WordPress agent as the example for the learnings and best practices below.
I created a dedicated folder for it and started with a simple prompt: “You are my website agent designed to help me manage my personal website.” Then it started asking me follow-up questions and we were on our way. No code required.
1. Pick the model and the effort
I experimented with different models and settled on Opus with High effort as the middle ground:
- Sonnet doesn’t work well as the orchestrator (the main agent that plans the work and hands parts of it to helper agents).
- Fable is overkill for what I need.
- For effort, Medium to High works well.
2. Give it a purpose first
In my experience, the most important thing for an agent is a clear purpose, and that’s the CLAUDE.md file. It’s a plain text file in the folder that the agent reads at the start of every session.
Once that was set, the next steps were:
- Building the agent’s memory, so it keeps learning.
- Teaching it skills and playbooks, for example how to run site operations and how to help me publish a post.
3. Manage the context window
As you keep chatting with the agent, the context window starts filling up. With Claude I get 1 million tokens, which sounds like a lot, but it eventually fills up.
My best practice is to keep a session state file (a short note of what’s in progress) and to regularly ask: “Update your skills and learnings so I can compact.” This way everything is saved, and I can compact the conversation when I choose.
4. Plan around the usage limits
My weekly usage limit resets on Sunday evenings, and you can only use so many tokens in a single 5-hour session.
So I now run scheduled routines to balance my usage through the day. These routines are automated sessions the agent starts on its own, for things like site maintenance.
I’m on the Max plan with 5x usage, and it’s the sweet spot for me with my current goals.
5. Clean up the instructions every month
Once a month, I do a prompt audit and ask the agent to improve and optimize itself. The agent and the way I use it keep evolving, so I want the instructions and skills to stay current, without garbage from old sessions or any conflicts.
6. When I get stuck, ask the agent
Often when I get stuck with the agent, I just ask the agent itself and have it give me three options.
What I’m experimenting with now
I’ve also been experimenting with spawning sub-agents that help complete tasks, with the main agent as the orchestrator for complex research tasks.
On top of that, I’ve been getting the main agent to talk to another agent I’ve been building as an executive assistant. These two are next level, and they still amaze me.
It keeps getting better
The more you use the agent, the more it learns, the more you can tweak it and the more it can help you. It’s a symbiotic relationship that keeps growing.
Overall, it’s been a fun experiment using an agent to help with my website and to keep growing my AI knowledge. It’s fun to talk to the agent in natural language, and to build in the guardrails and hard constraints that keep it safe. I wrote about the guardrails I set in How I used AI to improve my website.
How are you using your agents, and do you have any best practices? Let me know.