How I use my computer

The numbers

I started using Claude Code at the end of January 2026. The logs say:

• The March → April jump is when I stopped treating it as a chatbot and started handing it whole jobs
• 31% of my active time now has two or more sessions running in parallel — sounds chaotic, is actually fine (each project keeps its own context, see the vault page)

The vault

• One Obsidian vault of plain-text notes holds basically everything — every project, client, decision, invoice, half-thought. Git underneath, synced every 15 minutes, also reachable from my phone
• The unusual bit: it’s become Claude’s backend more than my notebook. There’s a filing map it follows, an index of every project, and a weekly lint job that flags drift. I mostly don’t open Obsidian — I just ask for things (‘pull up the Huw invoices’) and the right file comes back
• This is the thing that makes running several sessions at once workable — the context lives in files, not in any one conversation
• More → /stack/vault

Talking instead of typing

• A lot of my input is dictated — often voice notes on walks that get transcribed and turned into work
• I’ve ended up with a little recipe for delegating by ramble: outcome / done-when / your call / don’t touch / come back. Missing slots get filled with stated assumptions rather than questions back
• This page came from exactly that — a rambly voice note, plus an audit

The automation fleet

58 launchd jobs run on a schedule (the crontab is empty, it’s all launchd). The ones I’d actually miss:

• Business briefing — twice a day, reads my email, call notes and calendar across three Google accounts and regenerates one markdown file: who’s paid, who’s gone quiet, what needs attention
• Daily log — at 23:45 a job looks at what actually happened (git commits, new notes, tracked calls) and writes the day up as a diary entry. My journal had been dead since May; this hasn’t missed a day
• Money — bank feeds → a nightly rollup with net worth, burn and runway, plus a daily net-worth line that subtracts the corporation tax I owe (the honest number)
• Safety nets — a nightly job commits and pushes every repo I touched, Apple Notes migrate themselves into the vault, the Kobo syncs when I plug it in, and a weekly job audits the whole setup against its own documentation
• One-offs — during the heatwave a job checked air-conditioner stock twice a day and only pinged me on a real hit at a non-gouged price. Making a watcher for something is now about a 10-minute job

The receipt printer

• A Raspberry Pi drives an 80mm thermal receipt printer in the corner of the room, 24/7. Message a Telegram bot (or hit a hotkey) → a paper chit prints
• Every morning it prints my calendar, a workout summary, and the ‘ugh field’ — the tasks I’ve been avoiding, on paper, where I can’t alt-tab away from them. Net worth on Mondays and Thursdays
• It’s new (July 2026) but it’s stuck immediately — paper turns out to be a great anti-notification
• Full write-up → /stack/printer

Apps I’ve made for myself

Local-first apps, each built by describing them to Claude, installed like real apps, launched from Spotlight. The ones in daily-ish use:

• LargeBrain — a keyboard-first mind-mapper (I was using Scapple ~4h a day on client calls and wanted to stop reaching for the mouse, so I had a clone built with my own keybindings)
• CRM — my pipeline as a kanban over one JSON file
• Large Money — personal finance frontend over the money data above
• LargeCoder — renders my live Claude sessions as readable typeset pages (EB Garamond, basically a nice reading view for the terminal)
• Atlas — a map of what I actually know vs think I know, mined from my notes
• Client Sites — every client URL + passphrase behind one search box
• Stretch + Meditate — phone PWAs that replaced two subscription apps
• Plus this website, which is a fake desktop OS, because why not

Stuff I’ve built for clients

• Open Loops — a capture app for a client at 80,000 Hours: press a hotkey, type a thought, it lands timestamped in a shared Google Doc. In daily use
• Momentum sites — each co-thinking client gets a small personal cockpit site: what we’re working on, actions, session notes
• An accountability portal — daily proof-of-work submissions with civic-style receipts and a nightly verdict email (it has an API endpoint so his Claude can post to it, which I enjoy)
• Call primers — ten minutes after a client call, an automation reads the meeting notes and files a short primer into that week’s plan doc

The flow-state layer

The non-AI stuff. This is the part that makes a computer feel like a joy to use — nothing here is necessary, all of it is satisfying:

• AeroSpace — tiling window manager, every window keyboard-addressable. Claude parks files it opens for me in a staging workspace so nothing steals focus
• Vim keybindings — in the terminal and in Obsidian (yes, Vim mode in a notes app — it’s genuinely good for editing prose)
• Karabiner — caps lock is escape, and cmd-arrow does emacs line-editing in the terminal specifically
• Espanso — 63 text expansions: :da types today’s date, :hand types a whole ‘write me a handover doc’ prompt, :fab launches Claude on the big model
• Warp + a statusline — terminal tabs per project, with the model name colour-coded so I can tell at a glance which brain I’m talking to

Other hardware I like

• Kobo + reMarkable — a /bookify command sends any URL or file to both (EPUB to the Kobo, re-typeset PDF to the reMarkable) plus Apple Books
• A DJI mic — for the walk-dictation habit above
• An old ThinkPad running Arch — for writing, synced to the same git repos
• The Pi + thermal printer — see above

The graveyard

Stuff I tried that didn’t survive, which feels worth listing too:

• An n8n executive-assistant bot — Telegram, morning briefings from Gmail + Calendar, voice-note routing, evening recaps. Built over a weekend, worked, parked when the free trial expired (revival plan filed away, idk if I’ll bother — most of what it did got absorbed by the automations above)
• A task manager (Flow) — superseded by the printer’s ugh field
• A Socratic tutor (Dave) — got forked into two schema-therapy tools I do still use
• A commitment-guardian skill (Cate) — built it, smoke-tested it, never used it
• Two previous versions of this website, and roughly twenty one-off scripts that did their job once and retired