Week 1: The taste test is still human

I've run personal AI systems before, with several agents and a dashboard to bring it all together. They didn't last. The agents were too wordy. They were more generative than agentic, and not autonomous enough. In the end I stopped opening the dashboard.

So this week I started again. A lot has changed in how agentic AI works, and I wanted a fresh run at it. The big difference this time is that I'm doing it in public. This site, and these weeknotes, are where I share what I'm building as I build it.

This week in short

Built: This site, from first mockup to live, on EmDash and Cloudflare.

Broke: For about an hour, the live site showed the test version, and my login stopped working.

Learned: AI can build it, but the taste test is still human.

Next: Weekly posts running, then my first real agents for email, calendar and to-dos.

Built

The site started with a feeling. A few years ago I built a project around guilds of communities, and I loved its brand palette. So it's back: the purples and greens, and the guild language. Projects are "charters" here, and each week gets a log.

The fonts took longer. The first mockups used serif fonts, and they felt old-fashioned. The alternatives felt like every other site that someone spins out with Claude. As much as I use AI on most of my projects now, the taste test is still human.

That's also why the pill-shaped tags and the logo bar went. I see a lot of AI sites and a lot of AI content, and I didn't want this to feel the same. I wanted it to feel personal, not another cookie-cutter site. It's the same with WordPress. Lots of people use the same themes. What makes a site stand out is how you shape it to fit.

The site runs on EmDash, a new CMS that runs on Cloudflare. I've built plenty of sites before, but this is the first time I've used EmDash for real. Coming from a WordPress background, it's been a good experience so far.

By Sunday it had four charter pages, one for each of the projects I'm working on, and a Monday routine that turns my week into a brief for this weeknote.

Broke

On Saturday the site went live. A little later, a test deploy took over the real domain. For about an hour, grahamhoffman.co.uk showed the test version of the site, and my admin login kept failing. The AI found the cause, fixed it, and wrote the fix into its own deploy rules.

It didn't bother me much. It did show me that when AI does the work on its own, I'm not aware of every step it takes. I'm fine with that. You need some trust and confidence in what the AI is doing. It's a website, and AI can debug a website and fix it faster than a person can.

Where I draw the line is anything I can't take back. If the site already had lots of traffic, I'd be more careful. If it was sending an email, that can't be undone, so we check first.

Learned

I've used what I call compound learning since I first started with AI agents. Every time the AI makes a mistake, like getting the copy wrong or not finding the right report, it fixes it and writes down what it learned. Next time, it reads those notes first and updates its skills or its logic.

It's the same as working with a person. Making a mistake is okay. What matters is learning from it and not doing it again. The domain mix-up on Saturday is now a rule in the deploy skill, so it shouldn't happen twice.

This week I also joined the RoboNuggets premium course, after watching Jay's videos for a while. I'd done a lot of the tasks before, but picking up tips, tricks and new skills over the last three or four days has been really valuable. It's always good to see how other people solve the same problems in their own way.

The other big part of the week was the community. It doesn't matter how much AI is involved, the human connection still counts. Sharing progress, getting feedback and talking it through with other people was a big part of my week.

Next

By the end of this week, I want two posts a week running: this weeknote on Mondays, and a Thinking out loud post on Thursdays. The plan is to feed what I do each day into AI and pull out a thinking topic for the week.

Then the dashboard. My first real set of agents will cover email, calendar and my to-do list:

• A to-do agent that builds my list from meeting notes and discussions.
• A calendar agent that knows what my day looks like.
• An email agent that reviews my inbox.

It's also the first time I'll try a proper multi-agent setup, most likely Hermes Agent. That's the real challenge this week: going from one conversation with Claude to several agents working together.

If you want to try this

If you haven't played with AI yet, start by watching some videos. I can recommend RoboNuggets on YouTube, and their premium course if you want to go further. There are plenty of other YouTubers covering this now too. Find the one that clicks with you.

A lot of them focus on building an agency, and so does the course I'm on. Nothing stops you from pointing it at what matters to you. For me, that was a blog.

Most of all, start talking to the AI, get comfortable and trust the process. Nothing you do now can't be fixed, changed or updated later. What I'm doing today will probably change in a month, when my priorities or the tools change. What I learn in the next couple of months will build on what I learned this week.