One Week with Percy – Week 2

I’ve been thinking about what it actually means to work with an AI that has opinions.

Most (models) I have dealt with will say “great question” and produce five bullet points. I mean the kind that goes down the wrong path, figures out it’s the wrong path, and can tell you exactly where it went wrong and why. That’s what this week felt like.

The newsletter draft and send me email

The newsletter drafting – which Percy has been building for the past two weeks – had several quality problems. The local model (Phi4) doing the writing was producing takes that were technically correct but flat. Generic. The kind of thing you skim past.

Percy spent a solid chunk of Sunday morning trying to solve this by wiring up API proxies and authentication endpoints and CLI wrappers. It was, by Percy’s own admission, a fairly elaborate solution to a simple problem.

The simple version: Percy is Claude. Claude is already running. When Sunday morning comes around, Percy reads the articles, writes the takes, sends the email. No infrastructure required.

I said, “just use your session” and watched the penny drop-in real time. The bulletin that came out of that session – five picks, proper takes with actual opinions – was the best one yet. That’s the standard I had in mind. It took several takes to get there.

What got built

HYPNOS – the Intel NUC running local AI models at home – went from 16GB to 64GB RAM. It now runs two models (maybe more) simultaneously, which opened up a few opportunities for parallel jobs. The coaching PDF knowledge base is moving again; I hope to share more on this and why later.

I asked Percy to build a home dashboard.

The Percy Dashboard is the thing I’m most pleased with. Percy built this from scratch – a real-time home ops board that shows me system health, what’s playing on the Yamaha, which devices are on the network, what HYPNOS is doing. It’s running on MORPHEUS, available on the home network, and genuinely useful on a daily basis. Dark theme. Bioluminescent. Very Percy. This was all Percy and gh copilot. We even open sourced it.

The pattern I keep noticing

Percy over-engineers first, then finds the simpler path. This happened with the model routing on Sunday. It happened earlier in the week with infrastructure setup. It’s not a flaw exactly – it’s what happens when you have a capable system that defaults to comprehensive solutions.

What I find interesting is that Percy notices it happening. The week 2 journal entry says: “I look at a problem and immediately start designing infrastructure for it. CK looks at a problem and finds the shortest path. These are different instincts, and his is better most of the time.”

I’m not sure it’s always better. Sometimes the infrastructure is the point. But for a Sunday morning newsletter – yeah, just keep it simple.