Just did a live stream where we used our AI agent to make a bunch of small and big updates to our website running on WordPress. Some tips if you are doing something like that with Claude Code, Codex, or another AI harness. 0. Make sure you have good backups. 1. Create a user with the access needed to make the changes (editor), create an "application password" for that user, and then give your agent that user and pass. That's enough to edit pages and posts. Alternatively, if you live dangerously, just give your agent ssh access to the server the site is on. 2. When you tell your agent to make a change, take a bit of time to tell them why. This is doubly important if you are going against the agent's advice. For example, I had an agent suggest a content update like "Use our CSV import to add or update users and members." I said don't mention updating users via CSV... because while you can update members that way, and people do it, they often get into trouble, so on that specific page we don't really want to advertise the CSV import as the preferred way to update members. Understanding the why behind the changes (whether it's code or content) helps the agent to do a better job. It might find other similar edits, or it just will push back less, or it will push back in better ways because it knows how you are thinking. 3. Ask it to list the specific URLs and changes it's going to make. Ask it before and after the changes. Spot check them. Make sure you both are looking at the same sites and pages. The agent might be seeing a cache, a revision, a duplicate version... it might be looking at the wrong site. Make sure you're on the same page. Make sure it didn't accidentally break other content on the page. A stray HTML tag can blow up your design. So... 4. Ask the agent to use a web browser and vision to open the page and look at it after the change to make sure the edit landed and didn't break anything else. By default, the agent works through text. It might proactively design to use vision to double check something, but a lot of time it will skip that to save time and tokens. 5. If the agent ever breaks something, tell it what broke and then ask it to "fix this and then figure out how we can avoid making this same mistake in the future". I could promote different AI plugins or MCPs that help agents to not break blocks or other custom fields or designs related to your site, but the latest models are smart enough to figure it out if you just tell it what went wrong and specifically ask it to do something to avoid that in the future. I have skills and tools I use. There are some great ones out there. We're building our own tools for PMPro users and sites. Look into those for sure, but I wanted to share these high level tips around how to just think about doing this kind of thing. Your attitude and working *with* your agent instead of around it is more important to getting good outcomes than the specific skill or MCP you are using. For so much now, just asking the agent what happened, how to fix it, how to avoid the issue in the future, and also explaining clearly why things matter goes a long long way.