Built from bed · 2026-08-22
This website was made
lying down.
No laptop open, no IDE, no terminal in sight. Just a phone, a few sentences — some typed with one thumb, most of them spoken out loud using speech to text — and Claude doing the typing on the other end. Scaffolded, committed, pushed and deployed before the tea went cold.
How to do this yourself
You need a machine somewhere with Claude Code installed and logged in to GitHub and Vercel. Then the phone becomes the remote.
- 01
Get comfortable
Lights low, phone in hand. Open Claude Code on the phone and start a session against the laptop at home. No keyboard, no screen bigger than your palm.
- 02
Say what you want
Plain words, typed or spoken: "make a small Next.js app, commit to GitHub, deploy to Vercel." That's the whole brief. Claude checks the tooling, the accounts and the conventions before touching anything.
- 03
Scaffold
create-next-app, Tailwind, shadcn/ui, a dark-by-default theme, favicons, analytics. The boring-but-important shell gets done the same way every time.
- 04
Write the page
One warm front page — this one. Fonts chosen, copy written, spacing tuned. You read it back on the phone and ask for tweaks in a sentence.
- 05
Commit & push
"initial commit" on main, a dev branch for ongoing work, a private repo created with the gh CLI under the right account.
- 06
Ship
vercel link, connect the repo, deploy. A URL arrives in the chat. You tap it, it works, you put the phone down.
The honest bit
The phone — or the remote control, really, since it's just steering a session running on a machine at home — isn't the clever part. What makes this possible is everything that was put in place long before bedtime.
There's a set of written rules Claude reads at the start of every session: which package manager, which UI kit, which host, which branch is which, what every new app ships with on day one. There's a checklist for bootstrapping a project and another for the app shell — theme, navbar, favicons, analytics — so none of that gets reinvented or forgotten. There are templates for the project docs. There's a design guide so the result has a point of view instead of the default look.
Then there's the plumbing. Several GitHub accounts kept strictly separate, resolved automatically by folder, with a pre-push hook that refuses to let one identity leak into another's repo. Vercel linked to the right team. SSH keys, CLI logins — all sorted, all tested, all boring.
And permissions. An agent that can run commands on your machine needs to be trusted carefully, not blindly. Every allowed command was considered and granted deliberately — enough that it can scaffold, commit, push and deploy without stopping to ask at every step, but no more than that. Getting that balance right, and revisiting it as the workflow grew, took as much thought as anything else.
That took real effort: weeks of refining an agentic workflow, writing down preferences that used to live only in my head, and fixing the edges every time something went wrong. The payoff is that a one-line brief from a phone turns into a correctly built, correctly owned, deployed app — because the system already knows how I work.
So: the phone isn't the clever part. The preparation is.