As I write this, I have sixteen Claude threads open. And I’m talking to myself — or rather dictating to Claude through Wispr Flow, because lately my ideas are coming out faster than I can type them. Then I read the Claude Code origin story, and it offered a hint about how to address the pain points I’ve been feeling all week.
Silicon Valley origin stories
I’ve always been a bit enamored of Silicon Valley origin stories — two people in a garage, building something the rest of us can’t even imagine yet. Dell started that way and so did many others. The Making of Claude Code is the latest entry in that genre, and I loved it. There’s something about people on the bleeding edge of a new technology that has always piqued my interest. One of my favorites is still Tracy Kidder’s The Soul of a New Machine, written about the invention of the minicomputer. I love learning about people building something conceptually brand new.
I started with Perplexity because I liked how it wrote
I started with Perplexity in July 2024, because I liked how it did research and wrote back to me; in retrospect, it was probably Claude doing that writing all along, but I wasn’t really paying close attention to LLM names and versions at the time. Even though I helped ship a machine-learning solution in 2018, in 2024 I was still in the infancy of using AI for my own business, operating with one open chat window and posing one question at a time. It’s worth remembering how recent that was – measured in months, not years.
Right now I have sixteen plates spinning
I switched from Perplexity to Claude about two years later. Before the merge Anthropic just announced, that meant three active threads in Cowork and thirteen more running separately in the browser and no way to see across them. These days, with the two environments merged, I can at least see all of it in one place. But managing sixteen active threads is still a lot of plates to keep spinning. It feels as though in the blink of an eye, Claude has completely transformed how I work. Ideas are coming to me quickly, and I love that I can kick off a thread to retain those ideas, turn around and do something else, then come back and pick up threads where I left them.
My brain has always had more ideas than I could capture, and this is the first time I’ve experienced the tools keeping pace. A lot of that speed is due to Wispr Flow; dictating gets my ideas out nearly three times faster than what I can type. (That’s a post of its own, coming soon.)
The catch is that sixteen threads is a lot of plates to keep spinning. Even when I’ve taken the time to name every thread and keep my projects organized, I lose track of which one belongs to which train of thought, and if they need to be done in a certain sequence. What I am wishing for next is a bird’s-eye view of every plate I have in the air at once.
From doing the work to conducting it
If you read Tuesday’s post, the one thing still on my wishlist is an orchestrator. This is the thinking underneath that wish.
For most of my career, doing good work meant either doing the work myself or delegating it to a member of my team. What’s changed with Claude is that I can be much more hands on, and I’m keeping the plates spinning myself — starting threads, keeping them moving, and pulling the results together at the end. I’ve been building agent workflows in n8n lately, where one master agent runs a few sub-agents underneath it, each taking its own specialized piece of the job. That’s the concept I’m imagining for the future of Claude. I’m managing all these threads, and it would be ideal to orchestrate and sequence them from one central place.
The next level of abstraction
In the Claude origin story, Igor Kofman (a member of Anthropic’s technical staff) says that as the models and Claude Code keep improving, we’ll move to a new level of abstraction. Instead of managing a bunch of Claude threads yourself, you’ll be managing “the Claudes’ manager” — one layer removed from the agents themselves. I’m looking forward to that.
The layer I keep wishing existed
So what would that layer do, concretely? Let me use a real example from my own work.
When I do user research, the research itself happens with real people — conversations, interviews, noting key points to review. Then I bring what I’ve gathered into Cowork to aggregate it and turn it into personas, segmentation, user requirements, and the other deliverables that come out of good research. Once those files exist, I move them manually into the right directories in Claude Code so they can inform whatever I’m building. It’s not the end of the world, but I’m the one moving the output of each step into the next. I’ve reorganized my hard drive and my cloud storage two or three times already, trying to make my files easy for Claude, then Code, and now Cowork to consume.
Right now I have three environments — Claude Chat in the browser, Claude Code, and Cowork — and I move between them all day. Ideally the orchestration layer would sit over all three and manage those handoffs for me. Research wraps up, and the personas.md file lands where Code needs it. A long-running job finishes in the background, and instead of waiting for me to notice, it passes its output along to the next thread — or breaks it into pieces and hands each one off — while I’m busy executing on another idea.
There’s a simpler path, too — autonomous agents like Hermes that just execute a sequence of tasks on their own. I’ve been impressed watching what they can do, but it’s also one more thing to keep track of. So that too is a post for another day.
I spent years working on portals, and this reminds me of a pattern I saw repeatedly in that work. As an organization matures and its content grows, if nobody curates it, the usual fix is to build another layer of navigation on top of the prior ones. Yesterday’s top-level navigation becomes today’s second level, and everything underneath gets pushed down. Similarly, a master orchestrator is that next layer — not a replacement for the environments I already use, but a level above them that finally ties them together.
Every plate is tied to something I’m trying to accomplish
The part that matters most to me is bigger than any single feature. Every one of those threads is connected to something — a piece of my business strategy, a project I’m trying to progress, an outcome I’m working toward. The threads look scattered, but underneath they all ladder up to something I’m trying to accomplish.
What I really want is to be able to tell Claude that. I can already shape how it works through Skills and Instructions. I’d love to articulate the strategy the same way, so Claude could help me orchestrate and prioritize across all of it. It wouldn’t just run the threads, but it would understand what they’re in service of and help me decide what deserves my attention next.
For me, the payoff isn’t only speed. The time I don’t spend hunting across threads is time I get back for the work that needs me, especially time with clients and my own writing. That’s the whole reason I’m automating the administrative part of my work in the first place.
I’d love to hear how your work has changed
If you’re still working in one chat window, that’s where I was a few months ago, and the jump is closer than it looks. If you’re already juggling a dozen, you probably understand exactly what I’m wishing for. Either way, I’d love to know where you land on that curve.
This is the kind of thing I’ll be exploring in this series — the evolution of work as new LLM-integrated tools emerge, and sharing how it’s changing my own work. At some point I’d like to do a LinkedIn Live or a small workshop to show what I’m up to and explore if it could help others. If that sounds useful, drop me a note in the comments.
Leave a Reply