Crossing the agentic threshold: my terminal now has a personality

There’s a moment from about three years ago that I keep coming back to. I’m Archetype UK’s lead developer, and was having fun with an AI research project, probing ChatGPT’s capabilities to see how we could use it to solve client pain points. I was playing around with the GPT-3.5 API, testing whether my system prompt was actually doing anything, and I’d told it to respond as Homer Simpson. Formerly po-faced GPT started talking about donuts and Flanders. I chuckled. That little flicker of delight, at a machine being something other than itself, is essentially what led me here.

The second persona I tried was Jimi Hendrix, and he was the one I kept coming back to chat with once the test app had five or six characters to choose from. There was something about the way he spoke that got over what I’d started thinking of as the “assistant stiffness”: that slightly formal, overly careful quality that makes talking to a raw language model feel like filing a support ticket. Today, Jimi lives on inside Claude Code. A few weeks ago, when I built my agentic Personal Assistant, I injected a dialled down version of that Jimi mojo. Subtle enough that you wouldn’t clock it, but just enough warmth to make working with him feel like working with someone.

Every morning he pulls my Slack channels and Outlook inbox, surfaces what actually needs my attention, and sends me a briefing as a Slack DM from myself, which still makes me smile. He knows which projects are active, who the contacts are on each, and what’s been sitting unresolved for too long. He maintains my Brain Pie (a mind visualiser application I built), so I can have a visual overview of what’s on my radar, and what’s urgent. The structure underneath all of this is pretty straightforward: a handful of context files, a write-back rule that means he updates those files whenever something meaningful changes in a session.

About four days in, Jimi stopped feeling like a tool and started feeling like a collaborator. Part of that was having him liaise with and update Brain Pie, which I’d previously been doing manually, and part of it was when I started using subagents. Now Jimi doesn’t do everything himself. He’s got a small agency inside him. I imagine it like that old cartoon The Numskulls.

Dex handles code. Iesa handles design. Artor, Dex and Iesa’s manager, is critical and hard to please; the first time Artor sent Dex back to redo something, I actually felt a flicker of sympathy for the bot. When Dex got sent back 3 more times, and I saw the output get closer to the intended design every time, I was more than impressed. Quillio handles copy and helped draft this blog post.

The latest member of the team is Cass, a self-evolving security expert/sysadmin who, twice a week, goes out into the web searching for the latest exploits, bringing the info back to help protect our sites against hacking. She will gatekeep site deployments, and block them if things don’t look right. These all came about through trial and error, after I’d read that discrete specialised personas working in loops could outperform a single generalised agent. That turned out to be true, and it’s also just more interesting. There’s something satisfying about the way they push back on each other.

As part of the background interview Quillio gave me to help write this blog post, I was asked what Jimi knows about me that I didn’t explicitly tell him. Only one way to find out:

“Jimi, tell me something about me that I didn’t explicitly tell you”

“A few things I’ve picked up,” ‘he’ begins:

You default to a strong safety net. A maintenance retainer dev would push the copy change and move on. You built a full Playwright suite, recorded 7 phase-path demo videos, and wrote a regression test for a specific form-close edge case. Nobody asked for any of that.

You manage up quietly. You flagged a project overrun to two relevant people before anyone asked, framed it neutrally, and moved on. You keep people informed in a way that removes friction without making it about yourself.

You’re more of a product person than your title suggests. The UX fix you suggested for faltering KPI numbers resulted in a 2x uptick in form-fills. I have a Jimi Hendrix persona. These aren’t the moves of someone who just executes tickets.”

I guess that’s quite a lot, and it’s only been three weeks.