Building an AI Chief of Staff with Claude Code

WorkflowSettings

The Core Mental Model

The "AI Chief of Staff" framing positions Claude Code not as a developer tool but as an executive productivity layer — an agent that unifies your inboxes, manages your todo list while you sleep, pushes back on decisions, and aligns your time to your goals. A CEO at Ada (Mike Murchison) reports it roughly doubles productivity (from ceo ai chief of staff claude code). A non-programmer communications consultant (Jim Prosser) built a full system in 36 hours for $5–10/month that replaces $400–1,000/month of virtual assistant work (from jimprosser chief of staff claude).

The bigger win is not time savings — it's cognitive load reduction. Starting the day in "decision mode" instead of "gathering mode" eliminates low-grade anxiety about email and operational overhead (from jimprosser chief of staff claude). The mental model for async AI work is not "grind during gaps" but "design your day differently because the work runs without you" (from dispatch async day design).

The Architecture: Four Layers

The most important insight from jimprosser chief of staff claude is that the architecture matters more than the code. Each component must know the others exist — the email scanner produces metadata the morning sweep needs, the sweep assembles context packages subagents need, and the time-blocker reads all upstream output. Layered automation compounds: each piece assumes the others exist, and the whole becomes much more than the sum.

Layer 1: Overnight Inbox Scan Run a scheduled task (or macOS launchd job) overnight that scans email, Slack, and other inboxes. Classify everything into four categories (from jimprosser chief of staff claude):

Default to "prep" over "dispatch" when uncertain. The system never sends emails autonomously — only drafts them.

Layer 2: Morning Sweep Each morning, the sweep assembles context packages for the day's tasks and dispatches to six specialized subagents running in parallel: draft emails, update client files, schedule meetings, run research, etc. (from jimprosser chief of staff claude). A CPO at Pendo runs a single Claude Code command that plans his entire workday (from claude code daily workflow vibePM). Darrin Henein has a "staff" of always-on agents via launchd that produces a tight, organized daily brief by 9am (from always on agents launchd obsidian).

Layer 3: Meeting Intelligence Automate meeting prep as a pipeline (from meeting prep tool claude code):

cron trigger (Railway)
→ calendar API (Unipile) pulls today's Google Calendar
→ Exa + Perplexity research each attendee in parallel (LinkedIn, company sites)
→ Claude formats a clean dossier per meeting
→ Resend delivers to email with meeting title + time in subject

Runs at 7am daily; dossier lands in inbox before your first meeting. Separately, use Granola MCP to automatically enrich contact records from meeting transcripts — the AI extracts context from conversations and updates relationship records without manual entry (from ceo ai chief of staff claude code).

Layer 4: Time-Blocking & Weekly Accountability After the morning sweep, a time-blocking component reads all upstream output and generates a geographically-optimized calendar from remaining tasks (from jimprosser chief of staff claude). For memory, add one line to CLAUDE.md (from weekly recap agent memory):

When you complete work, log it to memory/weekly-recaps/current-week.md

Every task completion auto-updates the recap file — by Friday you have a full week log without writing a single thing yourself. Darrin Henein takes this further with a weekly AI coaching conversation that reviews meeting transcripts, task progress, and goal alignment, then delivers honest feedback on how he spent his time (from always on agents launchd obsidian).

The Critical Design Principles

From jimprosser chief of staff claude, the most important decisions are about where the human-AI boundary sits:

The system doesn't need to be perfect. The value is shifting your starting state: instead of spending the first hour gathering and triaging, you start reviewing and deciding.

The Technical Setup

Scheduling options (from jimprosser chief of staff claude and always on agents launchd obsidian):

MCPs to connect (from Settings):

CLAUDE.md configuration: write the system in Markdown files describing desired behavior, not code. Jim Prosser, a non-programmer, built his entire system by writing detailed markdown behavior descriptions and having Claude implement them, iterating against real task lists — no software engineering required (from jimprosser chief of staff claude).

Physical interface: A Stream Deck + Claude Code combination creates a one-button trigger for complex multi-agent orchestration (from jimprosser chief of staff claude).

What to Build First

Start with the single highest-leverage piece. From the real-world implementations:

  1. Start with the morning brief — launchd or scheduled task, produces a daily summary by 9am
  2. Add meeting prep next — 5-API pipeline, fully automatable, high daily value
  3. Add inbox triage — after you trust the brief and meeting prep, add the classification layer
  4. Add parallel subagents — once triage is stable, dispatch specific tasks to specialized agents
  5. Add weekly accountability — the one-liner CLAUDE.md addition for auto-logging

Each step is independently valuable and compounds with the others.

Sources cited: