How I use Claude

I’m not a programmer — Claude writes all the code here, and I’ve ended up using it for most knowledge-work too. The honest shape of that, from its own logs.

The ramp

Messages to Claude per month, 2026† month in progress471Feb2,995Mar35,995Apr31,257May54,460Jun25,340Jul77,805Aug18,916Sep
  • January was 62 messages of poking at it. The April jump is when I stopped asking questions and started handing over whole jobs — a goal and a done-condition instead of a prompt
  • June was 54,460 messages and ~16,000 tool calls

The setup

  • Skills140 installed skills and slash-commands: /money answers questions from my actual bank feeds, /inbox triages my messages, /grok prints quiz chits, /voice drafts things in my voice from a pack mined off my real sent email (it was used on this very page, with mixed first results 😅)
  • Memory186 small files of accumulated corrections and context. The rule I’m going for: never have to say the same thing twice
  • Hooks — code that runs around Claude: a session-start hook injects my formatting rules into every conversation, and a finished job pings a notification that deep-links back to the right terminal tab
  • The Freedom hook — killing my app blocker (which I do a lot) now triggers a mandatory 5-minute chimed pause before the command actually runs. Claude enforces a sit before I can doomscroll, which is uh, rather effective
  • Subagents1331 delegated tasks in the last 30 days. The audit behind this page was five of them sweeping the machine in parallel while a sixth checked their work
  • MCP — Claude can see my calendar (three Google accounts), time tracking, messages, workout log, and browser. ‘Can you check…’ is a complete instruction now

Talking to it

  • A lot of input is dictated, often on walks. The recipe I’ve landed on for a good brain-dump: outcome / done-when / your call / don’t touch / come back
  • Missing slots get filled with stated assumptions rather than clarifying questions, which is the difference between a 60-second reply and 10+ minutes of finished work
Photo: the ramble recipe card

Before all this

  • There’s a fossil record: for an executive-assistant application in 2025 I made a deck with a ‘productivity stack’ slide — Toggl, a text expander, Maccy, TextSniper, Freedom, Intend, Beeminder, Anki, Miro, Alfred, an early dictation tool
  • Good tools, but each one was something I operated by hand. Half of them are still here (Freedom, Anki, Obsidian, the dictation habit) — the difference now is the layer that operates things for me
Photo/embed: the ‘productivity stack’ slide from that deck

← the stack