Build an AI employee: the method

Overview

Every playbook in this section follows the same seven steps. Learn them once and you can build an agent for any role — the role-specific pages just fill in the brief, the questions, and the automations.

Step 1: Hire the agent in the right department

Departments keep files and connections separate, so put the agent where its work lives: a support agent in Customer support, a finance agent in Finance. Create the department if it doesn't exist, then click New employee (⌘D), give it a name and a one-line description, and let the workspace finish setting up. If you haven't yet, connect a model.

Pick a real name. You'll be talking to this agent every day, and "Maya" reads better in a WhatsApp thread than "support-bot".

Step 2: Write the role brief

Open the agent's settings → identity and write a first version of the brief. It doesn't have to be complete — the interview in step 3 will fill it out. It does have to answer four things:

You are {{agent name}}, the {{role}} at {{company}}.

## Who you are
One or two sentences: the company, what it does, and where you sit in it.

## What you own
- The two or three jobs you are responsible for.
- What you hand back to {{owner name}} and what you finish yourself.

## How you work
- Tone and length. Language. Formatting habits.
- Ask one clear question when something is missing rather than guessing.

## Never, without my approval
- Anything irreversible: sending to customers, paying, deleting, signing.
- Anything about money, legal terms, or personal data.
- If in doubt, draft it and ask.

Replace the {{…}} values and click Save. Each playbook has a complete, role-specific version of this brief.

Step 3: Let the agent interview you

This is the step most people skip, and it's the one that makes the difference. Instead of trying to write everything down, open a chat and have the agent ask you about its own job. It knows what a good support lead or finance manager needs to know; you know the answers.

Click the agent's card to open the chat. If you have a Business Context Profile, attach it with the paperclip first. Then paste:

You are my new {{role}} at {{company}}. Before you start, interview me about the business so you can do this job well.

Ask one question at a time and wait for my answer. Keep going until you have a clear picture of:
- what this job is, and what "done well" looks like
- who our customers are and what they usually need
- the tools, documents, and people involved in this work
- how I like to communicate and how detailed you should be
- what you can decide on your own, what needs my approval, and what you must never do

Start with the most important question. If I've attached a Business Context Profile, read it first and skip anything it already answers. When you have enough, tell me and we'll move to a summary.

Answer in plain language, as you would to a new hire. It usually takes 10–20 questions. The role pages list the questions a good agent will ask, so you can prepare answers in advance.

Step 4: Fold what it learned into the identity

When the agent says it has enough, ask it to write the brief for you:

Summarize everything you learned as an updated version of your identity: who you are, what you own, how you work, and what you must never do without my approval. Keep the structure of my current brief, write it in the second person ("You are…"), and keep it under 400 words. Facts that will change (prices, current projects) go in a separate list at the end, so I can put them in a file instead.

Read it. Correct anything it got wrong — this is where misunderstandings surface cheaply. Then copy the brief into identity and Save, and put the "facts that will change" list into a file (next step).

Step 5: Connect what the job needs

Now give the agent the same access a new hire would get on day one:

Start with what the first task needs, not everything it might one day need. Mention the connected apps in the identity so it knows they're part of the job.

Step 6: Schedule the recurring work

Anything the agent should do without being asked — the Monday report, the daily inbox summary — becomes an automation: a prompt, a schedule, and a channel to deliver to. Run the prompt by hand in chat first; when the output is what you want, paste the same prompt into automations → Create and set Send to to your WhatsApp or your team's Slack.

Each playbook lists two or three automations with the exact prompt and frequency.

Step 7: Review the first week, then refine

An AI employee is still an employee. For the first week:

• Read every conversation in Chat history, including the ones from WhatsApp or other channels.
• Check each automation's first output. Was it useful? Right length? Right channel?
• Note every correction you made twice. Each one is a missing line in the identity — add it.
• Note every question it asked that it should have known the answer to. Each one is a missing fact — add it to the file.
• Decide what it has earned: can it now send that email itself? Update the Never section accordingly.

After that, a monthly skim is enough. When you hire the next agent, Duplicate a good one and edit the brief — a well-built agent is the best template for the next.

Playbooks