100% AI — 10 days in

Ten days. That’s how long it’s been since the first post in this experiment — ten days of running everything I can through AI. Enough for an honest first check-in.

Day-10 takeaway: there is an upfront learning tax, but the quality upside is huge once the system starts compounding.

A rollercoaster, and the “upfront tax”

What I’m going through feels like the learning tax you pay on every new subject. Some tasks require a genuine paradigm shift, and I’m on a rollercoaster: there are moments when I think, “this task that used to eat my time — I’ll never do it by hand again, or almost never.” Then, a few days later, I hit the edge case that breaks it: I have to go back to the skill I’d configured and make it more versatile, because a rare situation keeps showing up. The tool is never “done” on the first try.

Plan more, produce better

The real shift is here: I spend a lot more time thinking and planning up front. For the projects I have in mind — DemoInk, which I’ll present soon, and others — planning clearly takes longer than before. But the result is immeasurably better than what I’d have produced normally. It’s counterintuitive: it feels like you “lose” time at the start, and you get it back a hundredfold in the quality of what comes out.

Connecting systems — and the session puzzle

A large part of my time goes into figuring out how to connect systems together. The MCP standard lets you build bridges between just about anything, and that’s pretty wild when you think about it. The real question remains: what memory do these systems actually share with each other?

Then there’s working in parallel. I try to keep several sessions running at once, so I can kick things off while the rest is working. The flip side: every morning I spend a good twenty minutes getting all my sessions back in order so the work can resume. It’s not hell, but at the slightest interruption I feel like I have to restart that “session page.”

What I’ve tested on the tooling side

• Claude Code in the terminal inside VS Code: fairly convincing.
• The Claude Code desktop app: I can’t find a strong use for it — I can’t really parallelize tasks. I wonder whether the “cowork” side might be the right way in, but for now, no.
• The Claude Code extension for VS Code: not bad, but a few details bug me — session and history management, and MCP permissions.

Where I’m at

Overall, it’s still a great experience, and genuinely interesting. I’ll admit there’s real peace of mind in no longer watching the tokens I consume, and being able to focus on results. I’m in a phase where a lot of ideas are moving in parallel: I’m optimizing how I work, new projects are starting, and I’ll soon be able to ship some things concretely — starting with the simplest ones.

---

Next milestone: turn this pace into consistent, concrete deliveries.

Stay tuned.