Poolside Director
I was standing in the boarding line at the airport, telling my phone what I wanted, and watching an AI configure itself to do it—without me touching a settings screen.
I'd set up OpenClaw a few days earlier on a Mac Mini at home. The idea was an always-on Claude instance I could reach via Telegram: persistent memory, cron jobs, its own accounts, its own credentials. A virtual employee who never clocks out. I treat it like a semi-trusted contractor—own Google account, own GitHub account, spend-limited API keys, no access to my personal email or calendar. When I want it to handle something, I forward it or send a Telegram link. It runs in a non-admin user account on the Mac Mini, sandboxed by the OS and the limited credentials I've shared.
I got it minimally working, then flew to Florida with my family for a week. I also jokingly (sort of) asked a couple of buddies to be on call to run to my house and unplug it if it went all Electric Dreams on me.
The plan was to check in occasionally. What actually happened was I started issuing directives from my phone while waiting to board. When I forward you an email, summarize it and create a todo. Set up a project for ideas I want to capture. If I send you a link, evaluate whether it's relevant to what I'm building. No config files. No YAML. Just... telling it what I wanted.
It worked. By the time we took off, the system was running.
The interesting part wasn't that it worked—it was how it worked. OpenClaw has a rich configuration UI. I never opened it. Every tweak, every refinement, every standing rule emerged from conversation. When something didn't behave right, I'd point it out over Telegram, and it would adjust. The feedback loop was entirely verbal. I was configuring enterprise-grade automation the way I'd ask a colleague to handle something.
By mid-week, I got ambitious. I had it clone one of my projects—a low-stakes greenfield game I've been noodling on—and told it to act as PM. Not just track tasks. Run the project. It spun up Claude Code instances using my personal workflow plugins, assigned work, reviewed PRs, filed issues. I was poolside in Florida, occasionally glancing at my phone, while an AI project manager ran an AI engineering team building software I'd designed.
I'm aware this sounds like a fever dream. It felt like one too.
The stakes were low—a hobby project, no production users, nothing breaks if it's wrong. But the shape of the workflow was real. I was a director, not an implementer. Not even a PM. I was the person who occasionally said "that's not quite right" and watched the system correct itself.
Here's where my brain went: I'm part of a tiny software startup. We're perpetually drowning in glue work—the operational overhead that isn't building product but makes building product possible. Task triage. Status updates. Making sure nothing falls through the cracks. What if we stood up an OpenClaw instance as a virtual team member? Not replacing anyone. Handling the stuff that currently doesn't get done because everyone's too busy with "real work."
I haven't pitched this yet. Not seriously at least. But I keep thinking about it. I've started drafting a real proposal.
Eleven days in, the thing that surprised me most wasn't the capability—it's that the capability composes. An assistant that handles email can also handle tasks. An assistant that handles tasks can become a PM. A PM can run engineers. Each layer builds on the last.
I'm back home now. The Mac Mini is humming beside me. Morning briefs show up at 8:30 AM. Forwarded emails get processed overnight. When I forget about something, it reminds me.
I'm not sure what to call this. But "assistant" feels too small.