Our youngest is six and deep in the phase where letters are magic. She fills pages with experiments: her name, words she’s sounded out, invented spellings that are almost right and somehow more charming for it. Reading is clicking into place in real time and it’s one of those things you can’t quite believe you’re watching.

We started doing word search puzzles together. She loves the hunt, scanning rows and circling letters the moment something clicks. Simple grids, short words, themes she picks herself (animals, food, family names). We got through the store-bought ones fast enough that it became clear we needed a way to make our own.

For a while I was doing what most people do: one AI assistant, one big context window, everything in one place. Convenient on paper. In practice, it’s a bit like asking the same person to be your chef, your accountant, and your personal trainer. The context bleed alone makes everything worse.

I’ve been running a different setup for a few months now and it’s holding up better than anything I’ve tried before. Three agents, each with a distinct identity and a narrow job. This is the writeup.

I self-host Planka on my homelab and use it to manage everything from software side projects to household finances to recurring chores. It’s open source, it’s fast, and it does what I need without the bloat.

What I wasn’t happy with was the friction of keeping it in sync with what I was actually doing. Every time I finished a task, I’d context-switch to a browser tab, find the card, drag it to the right column, maybe update the description. Thirty seconds per card, but across multiple boards and dozens of cards a week, it compounds into overhead rather than organization.

Every time I ask an LLM to draft something for me, there’s a version of the output that reads like it was written by a committee of consultants who all attended the same TED talk. You know the voice: punchy, agreeable, decorated with em dashes, and somehow both confident and saying nothing at all.

That’s the default. And for a while I accepted it, because editing AI output into something that sounds like me was still faster than writing from scratch. But it’s a tax. Every draft needs the same corrections: strip the buzzwords, collapse the dramatic one-liners, ground the abstractions in something real. At some point the pattern becomes hard to ignore.

Here’s a thought experiment I’ve been using in conversations lately.

Imagine you have unlimited resources. You can assemble any team you want and you have unlimited tokens. Would you launch a brand new product in two days, built entirely by AI?

Most people say yes, or at least a confident “probably.”

Now change the frame. You have a product with 500 million users, half of them paying. Clear market fit. Would you replace it tomorrow with something entirely built by AI?