My chief of staff, Claude Code

The dream of modern personalized technology has always been some kind of perfect digital assistant—a chief of staff that triages your inbox, organizes your day, and handles the operational overhead of your life before you've had your first coffee. For most people, that dream still lives in the vague future, somewhere between "Siri, but good" and science fiction. It's achievable today, even for people who aren't programmers. I know because I just built it over 36 hours with Claude Code. I should be specific about my coding credentials here: I have none. I'm a 43-year-old tech communications consultant who runs a solo practice in Marin County, California. But I designed this system, and the system works because Claude Code lets me think and then takes care of the rest. What It Actually Does Every morning before I wake up, two automated processes run on my always-on Mac Studio. The first scans the present and next day's Google Calendar for meetings at physical locations, calculates real drive times using the Google Maps API, and creates transit time events so I know when to leave and don't double-book anything. The second triages yesterday's email, identifies anything requiring action, checks my task manager (I use Todoist) for duplicates, and creates properly attributed tasks with priorities, due dates, and duration estimates. By 6:15 AM, my task system is current without me touching it. When I get to my desk, I hit a button on my Stream Deck labeled "AM Sweep." This triggers a command in Claude Code that pulls my tasks along with my calendar and recent meeting transcripts for context, then classifies every item into one of four categories. Green means Claude can complete the task fully on its own. Yellow means it can get 80% of the way there and I finish. Red means it needs my brain or my presence. Gray means it's not actionable today. I see the full categorized list, adjust anything that looks off, and say "go." Here's where it gets interesting. Six specialized AI agents fire in parallel, each with their own context window and scoped tool access. One drafts emails (it never sends, only drafts). Another updates my client files in Obsidian, my knowledge management system. Another schedules meetings. Another runs background research on a prospect, a topic, or a news event. They all work simultaneously while I focus on strategic work: writing, calls, the stuff that actually requires my brain and judgment. A couple minutes later, I get a completion report. Tasks marked done. Email drafts sitting in Gmail for my review. Client notes updated. Research filed. Next steps flagged. Then I press a second button on the Stream Deck called "Time Block" that executes another command that turns my remaining Todoist tasks into a time-blocked calendar using the duration estimates already assigned above. It knows which tasks need to happen at home versus the office versus specific locations. It batches all my errands into a single outing, routed geographically to minimize backtracking, with real Google Maps drive times. It schedules gym time on the right days. It puts home tasks in the evening window after my kids are in bed. If something doesn't fit today, it recommends a specific future day based on how loaded that day already is. I review the proposed schedule, adjust if needed, and say "go." Calendar events appear. My day is structured. One minute. The first time I watched the whole thing work, with six agents running in parallel, my terminal filling with progress updates, I just stared at the screen and laughed. Now I just jam to some music and start cracking away at tasks. The Architecture (Not the Code) The entire system runs on Claude Code. If you haven't used it: you describe what you want in plain English, and Claude writes and executes the code. The key capability that makes my system possible is subagents, basically independent AI workers that Claude Code spins up in parallel, each with their own instructions, tool access, and context. But the code isn't the interesting part. The interesting part is the design. The critical insight was figuring out the layering: each component had to know the others existed. The overnight email scanner doesn't just dump tasks into my task manager; it attributes them with the metadata that the Morning Sweep needs to classify them correctly. The AM Sweep doesn't just triage; it assembles context packages that each subagent needs to do its job in an independent context window. The time-blocker doesn't just schedule; it reads the output of everything upstream and accounts for it. You get the picture. This is what separates a system from a collection of scripts. Each piece is designed to feed the next one. Remove any layer and the others still work. But the whole is dramatically more than the sum: six independent workers holding different contexts simultaneously, none of them competing for attention or memory like talking to a chatbot. The actual building was a conversation. I wrote d