Operations
A week of running my business from a phone
My Mac was away for about a week in October, so everything went through my phone. What worked, what didn't, and the rule I ended up with about what an AI is allowed to ask me to do.
For about a week in October I didn't have my Mac. The outcome I'd written down on 7 October was blunt: the box can do everything the Mac does, because I have no Mac for about a week. Everything I did for my business that week, I did from my phone.
It mostly worked. I approved work, answered questions, sent emails, reviewed websites and kept an agent factory moving. It also showed me, quickly and repeatedly, every place where my setup quietly assumed I was sat at a desk. This is what I learned.
The set-up
The principle that makes this possible is simple. The phone doesn't do the work. Something always on does the work, and the phone is a window onto it.
The always-on part is a small rented server I call the box, which I wrote about in Building the box. It runs my AI sessions, answers my WhatsApp groups and keeps my files in step with GitHub every 10 minutes.
The windows are five things on my phone:
- The Claude app. The box shows up under Devices, next to the Mac. I open a session running there and type to it as I would at a desk.
- Orca Mobile. A phone app for the Orca workbench, connected over a private network between my devices. It gives me chat tabs and, when I really need one, a terminal. More in Using Orca.
- WhatsApp. My Claude groups answer messages, voice notes and photos. It's the quickest way in for anything small. I wrote about that in WhatsApp is how I talk to my AI.
- The reviewsheet. I type one word and get one link to a phone-friendly page listing what was built, the decisions I need to make with a recommended answer for each, and every open ticket in priority order. See The reviewsheet.
- Dev-site links. Every change to my websites goes to a private test version first, so I can tap through it on my phone before anything goes live. See The dev site is the thing I test.
Emails still go out from me. The AI drafts, I read, I send from the phone. That's deliberate and I'd keep it that way anywhere.
What the numbers say
The box saves its work to git (the system that records every change) with a "box:" label, so I can see how hard it worked. In late September and early October it was saving between one and eight times a day. On 7 October it was 17, then 23. From 9 October it ran between 35 and 57 a day. That's the Mac's share of the work moving across, and it's roughly what I'd expect.
What that number doesn't show is my time. Some days that week were quick. On 12 to 13 October I spent hours in chat, and that was the clearest lesson of the week.
What went wrong: rich text
The first friction was copying. I'd asked for a formatted email I could paste into the Outlook app on my phone. On 7 October it pasted as flat text: no numbers, no bullets, no bold, blank lines gone. My reply at the time was "Well that didn't work?"
The quick fix was to put the structure into the characters themselves: literal "1)", bullet characters and blank lines, which any app keeps. I sent that version and it was fine.
Then the AI tried to fix rich copy properly, twice, and both failed on my phone. The real reason was a rule in the phone's browser: it only puts formatted text on the clipboard from a secure page, and the page it had built wasn't one. No amount of rebuilding the page could have worked. The fix, once someone looked up the constraint first, was to serve the page securely. The failure note says it plainly: anything that depends on what my phone's browser can do has to state the limit and how it was established before building, and say it's untested on my device.
What went wrong: steps I couldn't take
The second friction was worse, because it kept happening. Three times on 9 October an AI handed me something I couldn't do from a phone.
It asked me to run admin commands on the server when I had no terminal access in that session. My answer: "How can I run any of those sudo commands without terminal access". It asked me to edit a ticket card by hand. My answer: "I can't edit the cards, i only interface with the mfo" (the orchestrator, the AI session that runs my factory). Both were solved once I pushed back: one needed no admin rights at all if done a different way, and the AI could edit the card itself.
The same day I tried to run the one admin command the server needed through Orca on my phone. It asked for a password that doesn't exist. Admin access had only ever been set up from the Mac, and the route agreed for doing that one command from anywhere had never been installed. It was fixed that evening.
On 12 and 13 October it happened again in a different form. A sign-in for a new account was sent to me three times with steps that needed arrow keys, which the Orca phone app can't send. A step that asked me to paste a key failed and there was no fallback.
The root cause was the same each time. The AI reasoned from what it was blocked on, not from what I could actually do. A missing permission on its side became a job for me.
What went wrong: too much chat
The 12 to 13 October sessions taught me one more thing. I spent hours in long, multi-topic chat messages, going back and forth, when the decisions could have been laid out one at a time on the review page with a recommended answer for each.
On a desk, a long chat is tiring. On a phone it's much worse. You lose your place, you scroll past the one question that mattered, and you answer something you half read. A day earlier, on 11 October, a list of items the review page was skipping had quietly hidden some stuck website tickets from me for over a day, and I'd had to chase them. Both point the same way: on a phone, what isn't on the one page doesn't really exist.
Links I couldn't open
One older lesson came back that week too. Back on 17 September I'd asked AI sessions to stop publishing pages as links on a service I couldn't open when I switched between accounts on my phone. Internal pages now go to a plain file on the machine, or to a page served privately to my devices. That week it saved me a lot of dead links.
The rule I ended up with
After the 9 October failures the rule got written down: before asking me to do anything, check it's something I can do through chat alone. I have no terminal by default and I don't edit cards or code. If a route is blocked for the AI, it exhausts the alternatives first, and only then asks, saying why each alternative failed.
It's a simple rule and it's the most useful thing to come out of the week. It applies to anyone you hand work to, not just an AI: if the person can't act on the request from where they are, you've moved the problem to them rather than delegated it.
If you want to try this
- Put the work on something that's always on. A phone is a window. Something has to be switched on behind it.
- Pick a few windows and know what each is for. For me, WhatsApp for small things, the Claude app for sessions, Orca when I need a terminal, one review page for decisions.
- Keep sending in your own hands. Let the AI draft. You send.
- Batch decisions onto one page. Each with a recommended answer, so a blank reply means yes.
- Tell the AI what you can't do. No terminal, no arrow keys, no desktop apps. Make it a rule, not a reminder.
- Test copy and paste into the real app. Formatting that works on a desktop often doesn't survive a phone.
- Set up admin access before you leave the desk. And check it works from the phone while you still have the laptop.
Would I do it again?
Yes, and I'll have to. In November I'm away for several weeks across a time difference, with the same set-up. The week without the Mac was a good dress rehearsal. Reading, deciding and approving all work well from a phone. Anything that assumes a keyboard doesn't, and the fix is to stop it reaching me in the first place.
If you'd like to work more of your business from your phone and want to know where to start, let's have a chat.