Fresh coat of paint

:tmil x Claude Design

The original design of this blog was, to put it nicely, boring. High contrast, no personality, nothing that made it feel like mine. I wanted something more appealing and easier to read, but I also wanted an excuse to push Claude Design beyond the small suggestions I had used it for before. This was my first attempt at a full redesign with it.

The process had three steps. First, I opened a normal Claude session and worked through what I didn't like about the old design and what I wanted to achieve. No design guidelines, colors, wireframes, typography... Nothing of that, just a conversation. The goal of that session wasn't to design anything, it was to generate a clear, specific prompt I could hand off to Claude Design. Once I had that prompt, I let Claude Design make the actual design decisions.

The result? Generic. My original prompt asked for something inspired by the asymmetric, unconventional layouts you see winning awards on Awwwards. What I got was another clean, readable design — nothing that surprised me or pushed in that direction at all. That's when I decided to spent a couple of days manually proposing the things the tool couldn't: the unaligned paragraphs, the page transitions, the reading progress bar, and a few other details.

Claude Design is genuinely useful for structured design decisions, but it's not yet a creative partner in the full sense. It executes well inside familiar patterns. The moment you push toward something unconventional, you're on your own and the tool becomes a collaborator rather than a creative lead. The creative direction still has to come from a designer.

Should I have hired a designer instead? For this project, no. The goal was to ship a better design, yes, but I also wanted to test the boundaries of what Claude Design could do. Now I know, if I had wanted something truly distinctive, then yes, I should have: there's a ceiling here that, for now, only a human designer can break through.

I need(ed) to get my finances in order

I have 8 bank accounts across two banks. For a long time, keeping them reconciled and tracked in one place was one of those tasks that lived permanently on my to-do list, but never urgent enough to actually sit down and do. The manual work involved was just too discouraging to start.

This month I finally did it. I used Claude and the WalletApp MCP server to plan and orchestrate the process, Python scripts to parse and reconcile months of bank statements, and a Playwright automation to drive WalletApp's web UI and import hundreds of payments and transactions, record by record. The app has no bulk import API and I was not gonna do it manually, so a bot clicking through the browser was the only option.

It worked. Mostly. Twenty-one records failed on the first run due to mismatched dropdown values. A post-import audit via MCP surfaced two phantom records invisible in the UI. One entry silently failed to save without throwing an error. There's also a small money gap between my computed balance and what the bank app shows. And I also caught a record I had manually entered wrong at some point with one digit off, and I had nearly logged 33 times the amount I actually transferred.

An automated script doesn't make mistakes: same input, same output, every time. AI and humans don't have that guarantee. That's not a reason to remove humans from the loop but instead it's a reason to be intentional about which parts of the loop actually need one.

This is still a WIP project, I'll write a much more detailed post once everything is fully closed and ready for future use. There's a lot more to say about the process, the tools, and what I'd do differently if I had to start over.