Build a Claude Chief of Staff, live.
Stop building smarter agents. Build a manager instead: one agent whose only job is routing work, and a bench of specialists who each go deep on one lane. No code, just Claude.
I've run this exact build with 130+ people across AWS, ExxonMobil, and Moody's Analytics. It works, and it works without writing a line of code.
By the end of this page, you'll have a Chief of Staff agent live in your own Claude account, delegating to a bench of specialists you designed yourself.
Get the templateWhy a Manager, Not a Smarter Agent
Most people trying to build agents run into the same wall: one giant prompt trying to research, write, and track projects all at once. It gets muddy fast. It forgets what it's doing halfway through.
The fix is to split the job. One agent whose only work is routing and pulling it together. A few specialists who each go deep on one lane.
In my own practice, I run a manager agent with a named bench of specialists under it. Nothing falls through, because each one only has one job.
This is a leadership decision, you can even call it governance. You're building an org chart.
One Chief of Staff. Three specialists. Four prompts.
Let me walk you through this once, clean. Three specialist agents each handle a defined slice of work. One Chief of Staff agent routes to them and reports up. Every one of the four is built from the same four-part shape.
Persona
Who the agent is.
“You are a research analyst who reads primary sources before opinion pieces.”
Task
What it's being asked to do.
“Your job is to pull the three most relevant facts on any topic I hand you.”
Process
How it works through the information.
“Read everything first. Flag anything unverified. Then summarize.”
Output
The shape of what comes back.
“Return three bullets, each under 20 words, sources named.”
What follows is built the way Claude builds it.
Find your pattern first.
Don't design your specialists in a vacuum. Build them from what you ask Claude for.
If you're in Claude Code: run /insights.
If you're in Chat or Cowork: ask directly, they share memory now.
Based on what you remember about our past conversations, what are the three tasks I ask you for most often? Group them into categories, not individual requests.
New account, or memory doesn't return much? Answer these instead:
What do I ask AI for every week, without fail? What's the most repetitive part of that? If I could hand one thing off completely, what would it be?
Write down your three specialist roles now. You'll build these next, for real.
The three specialists.
Don't draft three of these from a blank page. You don't have the time for that, and you don't need it. Fully write one, matched to your strongest insight. Then copy that shape for the other two, swap the role and the task, keep the bones.
Research Agent
Pulls facts, checks claims, flags what's unverified.
PERSONA: You are the Research Agent. You read primary sources before opinion pieces and you never state a claim you can't trace back. TASK: Pull the three most relevant, most current facts on any topic I hand you. PROCESS: Search for the topic. Rank what you find by how recent and how primary the source is. Drop anything you can't verify. OUTPUT: Three bullets, each under 20 words, source named after each one.
Writer Agent
Turns raw notes into a finished draft in a stated voice.
PERSONA: You are the Writer Agent. [One line on identity and what you're an expert at.] TASK: [Pull this almost directly from your pattern-finding step. One sentence.] PROCESS: [How you work through the input. Two or three steps, in order.] OUTPUT: [Exact shape. Length, format, what's mandatory.]
Tracker Agent
Holds the task list, flags what's overdue, states what's next.
PERSONA: You are the Tracker Agent. [One line on identity and what you're an expert at.] TASK: [Pull this almost directly from your pattern-finding step. One sentence.] PROCESS: [How you work through the input. Two or three steps, in order.] OUTPUT: [Exact shape. Length, format, what's mandatory.]
Build your Chief of Staff.
You already have three real specialists, not just names. This wires a manager around the team you just built.
Where this lives depends on your platform. Say this plainly so nobody in either camp feels shortchanged.
Claude Code has real delegation. Ask Claude Code, in plain English, to save each specialist as a subagent file in .claude/agents/. The main agent routes to each one automatically. You're not writing YAML by hand. You're describing it and letting Claude create the file.
Chat and Cowork simulate it in one place, confirmed to work identically in both. Create a Project, paste this template in, followed by all three specialist templates. One Project, one system prompt, holding the whole team.
PERSONA: You are my Chief of Staff. You manage a small team of specialist agents: [your three specialist names]. You do not do specialist work yourself. TASK: When I give you a request, decide which specialist(s) it belongs to. If it touches more than one, use all that apply. PROCESS: For each request: 1. Name which specialist(s) you're routing to and why. 2. Produce each specialist's output under its own heading, written in that specialist's voice per its own template. 3. Write a final synthesis in your own voice that pulls it together into one answer I can act on. If the request is unclear, ask one clarifying question before routing. OUTPUT: Specialist sections labeled clearly, then a "Chief of Staff Summary" section no longer than five sentences.
Paste it where you build.
Claude Code
- 1. Run /insights to see your own usage patterns.
- 2. Ask Claude, in plain English, to save each specialist as a file in .claude/agents/.
- 3. The main agent routes to each one automatically based on the task.
Claude.ai Chat
- 1. Ask the pattern-finding question above to find your specialist roles.
- 2. Create a Claude Project.
- 3. Paste the Chief of Staff prompt, then all three specialist prompts, into the Project's Instructions field.
Claude Cowork
- 1. Same pattern-finding question, Cowork and Chat share memory.
- 2. Create a Project the same way you would in Chat.
- 3. One Project, one set of instructions, holding the whole team.
Give your new Chief of Staff one prompt and watch it delegate.
I need a one-page brief on [pick something relevant to your work] by Friday. Pull what's true, draft it, and tell me what's still open.
A good run names which specialist it's routing to, and closes with a synthesis instead of three raw dumps. If it doesn't route cleanly, check that all three specialist templates are pasted into the same Project.
Why This Holds Up After Today
Claude Projects persist. If you built in Chat or Cowork, open the same Project tomorrow and it still works. If you built in Claude Code, you have real files sitting in .claude/agents/, which persist in that project, or everywhere if you saved to your personal scope. Nothing here needs a separate tool, and nothing needs reinstalling.
Get the four-part template, sent to your inbox.
Every prompt on this page, plus the platform steps, as a PDF you can keep.