How to Build a Personal AI Chief of Staff

Difficulty: Intermediate Time to set up: 2–3 hours Payoff: 5–10 hours saved per week

A Chief of Staff's job is to protect your time, surface what matters, and handle everything that doesn't need you specifically. This guide walks you through building an AI agent that does exactly that.

What Your AI Chief of Staff Will Do

• Triage your inbox every morning and surface only the emails that need your eyes
• Scan your calendar and prep a daily briefing (who you're meeting, what you need to know)
• Draft responses to routine emails
• Summarize long documents, threads, and reports before you read them
• Track open loops — things you said you'd follow up on but haven't yet

This isn't one agent. It's a small team of agents, each with a clear job. Here's how to build it.

Step 1: Define the Job Description

Before you write a single prompt, write out what your CoS should handle vs. escalate.

Handle autonomously:

• Scheduling acknowledgments ("Yes, Tuesday works")
• Summarizing meeting notes
• Drafting replies to subscriber/customer questions
• Morning briefing generation
• Flagging time-sensitive emails

Escalate to you:

• Anything involving money over $X
• Partnership or business development conversations
• Personal relationships
• Anything where being wrong would cause real damage

Put this in a SOUL.md file for your agent. It's the constitution that governs every decision.

Step 2: Set Up the Inbox Triage Agent

Tools needed: Email access (IMAP or API), an LLM with a long context window

The core prompt pattern:

You are an inbox triage agent. Your job is to:
1. Read the last 24 hours of email
2. Assign each email one of: [URGENT / NEEDS_REPLY / FYI / IGNORE]
3. For NEEDS_REPLY: draft a reply in my voice
4. For URGENT: surface immediately with a one-line summary
5. For FYI and IGNORE: batch into a digest

My email tone: [describe your style — e.g., "direct, no fluff, first-name basis"]
Things that are always URGENT: [list them]
Things that are always IGNORE: [newsletters, automated notifications, etc.]

Run this on a schedule — every morning at 7am works well for most people.

The key insight: The triage categories are more important than the drafts. Get the classification right first. Drafts can be improved over time.

Step 3: Build the Daily Briefing

Your briefing agent runs every morning and produces a single document:

# Daily Briefing — [Date]

## 🗓️ Today's Schedule
[Pull from calendar API]

## 📧 Inbox Summary
- X emails need your reply
- X urgent items (see attached)
- Digest of FYI emails

## 🔄 Open Loops
Things you said you'd follow up on:
- [Item 1] — last mentioned [date]
- [Item 2] — last mentioned [date]

## 📰 Relevant News
[Optional: anything in your industry worth knowing]

The open loops section is the most underrated part. Most people forget 40% of the things they commit to verbally. Your agent should remember what you forget.

Step 4: Memory and Context

Your CoS is only as good as its memory. Here's how to structure it:

Short-term memory (daily): memory/YYYY-MM-DD.md Log every email sent, decision made, and commitment given. Raw notes.

Long-term memory: MEMORY.md Distilled facts: key contacts and their context, ongoing projects, preferences you've expressed, things you hate dealing with.

Example long-term memory entry:

- David Chen (Acme Corp): evaluating our enterprise plan, price-sensitive, 
  deadline is end of Q1. Last contact: Feb 28.
- I prefer to respond to partnership emails with a call, not email.
- Don't book meetings before 9am or after 4pm.

Feed this context file to your agent at the start of every session.

Step 5: The Weekly Review Loop

Once a week, your CoS should generate a summary:

• How many emails were triaged
• What open loops are still open
• What recurring questions keep showing up (these should become templates)
• What took the most of your time (these are candidates for further automation)

This turns your agent from a static tool into a system that gets better over time.

Common Mistakes

Mistake 1: Automating without a feedback loop If your agent drafts emails and you never correct them, it'll never improve. Review drafts, edit them, and periodically update your SOUL.md with what you've learned.

Mistake 2: Trusting the agent with too much, too soon Start with read-only access (briefings, summaries). Add write access (drafting, scheduling) only after you trust the output quality.

Mistake 3: Vague escalation rules "Escalate important things" is useless. Define importance with specifics: dollar thresholds, person names, topic categories.

Mistake 4: No SOUL.md Every agent needs a constitution. Without one, it'll make inconsistent decisions and you'll spend more time correcting it than it saves you.

What This Is Worth

A good executive assistant costs $50–80K/year and handles maybe 60% of what's described here. A well-built AI CoS handles 80% for the cost of an API subscription.

The other 20% — the nuanced judgment calls, the relationship management, the things that require actually knowing someone — that's still yours. The point is to make sure you're only spending time on that 20%.

Next Steps

• Library subscribers: The AI CoS config template is in the Library under configs/chief-of-staff/
• Workshop subscribers: Bring your specific setup to office hours — we'll tune the prompts together
• Questions? Drop them in #support