Do What You Can't: One Weekend, Four Web Projects, Zero Excuses

Multiple web projects, neglected for years. Bad design, default WordPress themes nobody tailored, Docker containers running on vibes and prayers, one project not even started, and my own blog sitting there with a single post — over a year old. My coding background? Scratch, Commodore 64 BASIC, and whatever scripting the pro apps at work call “something with Studio”.

The challenge: use AI for one weekend to fix them all. Not just fix — transform. Link data types together: events, locations, blog posts. Integrate free public APIs. Responsive design. An AI chatbot. No bugs. Self-maintaining. Just a small nudge from me once a month — or whenever life happens.

Simple, right?

The Ingredients

  • Claude, cheapest subscription
  • A dream called Hermes — a future where I tell an agent via Telegram what I want, and it drafts proposals, I just approve
  • One fresh VM, new Debian install
  • First problem: no sudo. Claude suggests su. Crisis averted before coffee got cold.

Install Claude CLI. Discuss CMS options — needs to be free, needs an API Hermes can use to create and update posts, one database for all four projects, content shareable across them. Claude recommends Payload CMS. So Payload it is. Claude and I are becoming best friends.

I write prompts in Claude Web, paste them into Claude CLI in the terminal. Why can’t they just talk to each other? A question for another weekend.

Pages Taking Shape

Things are actually looking nice. Features landing. Then: bugs. Of course there are bugs.

Time to migrate content. Slightly terrifying: I give Claude direct access to my Docker VM, and it proceeds to break into the WordPress database, extract everything, and migrate it. Professionally. Respectfully. Without judgment about the state of that WordPress install.

Wake Up, Hermes

Content is in. Time to bring Hermes online to improve articles while I build new features with Claude.

First question: cheap LLM or expensive LLM for Hermes?

  • Cheap LLM: Burns through $20 fast, without building a proper connector
  • Expensive LLM: Burns through $20 slightly less fast, claims to have good connectors — but by the time it’s done claiming, the budget is gone

Either way: $0 left. No Telegram agent. No monthly content proposals. No “Hey, here’s a cool idea for a post.” The dream of a self-improving, self-maintaining AI agent quietly goes back to sleep.

The Bug Report (Honest Edition)

  • One project page has an unfinished layout
  • Claude credits nearly gone after two days
  • Can no longer save content in Payload directly — somewhere during migration, or possibly before, that broke
  • Not actually a problem, because Claude can write directly to the database. Wait. Wasn’t that Hermes’s job?
  • Timezone bug: Somehow I gave the server a US timezone on creation. Click, click, allegedly fixed. Still wrong. Claude insists everything is fine. The tides disagree.
  • Privacy policy: AI-generated. Half Danish context, half German content. A beautiful mess that somehow feels very on-brand.

Conclusion

For a Do What You Can’t project, I got surprisingly far. It felt like prototyping on steroids.

And when I burned through my Claude credits? Forced socializing with wife and family. Some house work, some garden work. Not the worst outcome.

Should I publish everything — including the bugs?

Why not. It was real.


← All posts