Derek Featherstone

After about 5 hours running my personal OS with Opencode and Sol 6.1 as the LLM, I have come to some conclusions: - nothing is load-bearing or doing the heavy lifting - nothing is worth noting - nothing is actually different

One of the best things I’ve read or heard this year was Peter Bell saying in a LinkedIn live something along the lines of: “Use LLMs to write code to replace themselves”

Working with AI → the rate of change is immense, and keeping up is difficult. One of the key tenets of the systems I'm building is resilience to change: key components being "swappable". Like the LLM. Or a CRM. Or an automated accessibility checker. Or... well, you get the idea.

I decided to see if I can get my Claude Code/Opus/Obsidian Vault setup running on OpenCode/Sol 6.1/Obsidian Vault. Once I get that running, I'll aim for the ability to use local inference rather than a cloud service. So far, so good.

Working with AI → the rate of change is immense, and keeping up is difficult. One of the key tenets of the systems I'm building is resilience to change: key components being "swappable". Like the LLM. Or a CRM. Or an automated accessibility checker. Or... well, you get the idea.

ENERGIZED!! That's how I felt after speaking at an internal conference on Thursday about “Accessibility & Inclusion in the Age of the Builder.” Wow, I'd forgotten how much I love presenting at events. Yes, it was virtual. That's not my favourite, but still love connecting over the internets.

Here's the latest approach I've been taking to autonomous work that gets authorized and queued up for my agents to work on.

My goals? Increase the portion that can run autonomously over time. Increase parallelization. Just getting started, but really liking how this is going so far.

Here's the latest approach I've been taking to autonomous work that gets authorized and queued up for my agents to work on.

And... the streak ended because two independent classifier agents disagreed on a result, so the Chief of Staff flagged it for my review.

Here's the latest approach I've been taking to autonomous work that gets authorized and queued up for my agents to work on.

- one plan went from 7 steps to 26 (from the adversarial reviews and fixes) and then the simplify skill brought it back down to 17 - 54 of the 118 steps that finished this week ran completely on their own, end to end.

Here's the latest approach I've been taking to autonomous work that gets authorized and queued up for my agents to work on.

If something significant changes, the plan AUTOMATICALLY gets marked as out of alignment... it comes back to me -- exactly how I'd want it. And if its all aligned, and agreed, the Chief of Staff kicks off the work.

Here's the latest approach I've been taking to autonomous work that gets authorized and queued up for my agents to work on.

3. Then, an independent reviewer agent checks the plan and decomposition without access to the model's reasoning (so it isn't just agreeing with what it previous decided!).

Here's the latest approach I've been taking to autonomous work that gets authorized and queued up for my agents to work on.

The goal is to get broad alignment on what we’re aiming for. And, if the plan isn't marked as aligned, the next step can't run and we can’t move forward.

Here's the latest approach I've been taking to autonomous work that gets authorized and queued up for my agents to work on.

Control your emotions, control your future. I apparently failed to control my future this morning, because I absolutely DID NOT control my emotions.

So... now... I'm home. I'm calm. I'm typing things to my agents, and I'm much more civil. There's a lot to unpack here in terms of understanding how we best design agentic experiences, particularly with a voice-based UI.

Control your emotions, control your future. I apparently failed to control my future this morning, because I absolutely DID NOT control my emotions.

After about 10 times, I lost it. Truly, I have no idea if it was 10 times... it honestly felt like 100. I'll check the transcript to find out. I was literally yelling at iOS Claude over carplay.

Control your emotions, control your future. I apparently failed to control my future this morning, because I absolutely DID NOT control my emotions.

... and it then speaks for 15 seconds, and then asks a question that I was in the middle of answering before it interrupted me.

Control your emotions, control your future. I apparently failed to control my future this morning, because I absolutely DID NOT control my emotions.

I've also created a simple message bus that goes between Claude on iOS and my laptop at home so I can ask things of my Chief of Staff agent. It's a really great setup for accelerating what I'm working on, and taking advantage of the drive time.

Control your emotions, control your future. I apparently failed to control my future this morning, because I absolutely DID NOT control my emotions.

Designing for extremes makes experiences better for everyone. I have Claude Code setup to connect via remote control by default. Then I can check in on my terminal sessions while on my phone.

When you consider an "extreme" case, you often find a better solution for everyone. And it's relatively easy to make that persist as well, with the caveats that memory entries and instructions files are really all just suggestions anyway :)

Designing for extremes makes experiences better for everyone. I have Claude Code setup to connect via remote control by default. Then I can check in on my terminal sessions while on my phone.

So... I've started sending this message: "For the next hour, I'm on via remote control on my phone so long replies are impossible to consume, and I've got my X yr old with me to bring me your messages and to reply to you.

Designing for extremes makes experiences better for everyone. I have Claude Code setup to connect via remote control by default. Then I can check in on my terminal sessions while on my phone.