How to Build AI Agents with ChatGPT Work and Claude Cowork

In my previous article, I introduced the 4 C's Framework:

Coordination. Creativity. Clarity. Coaching.

The idea is simple.

Instead of thinking about AI as one chatbot that you occasionally ask questions, think of it as a collection of specialised assistants that can help with different parts of your working day.

But there's an obvious question:

How do you actually do it?

The good news is that you don't necessarily need to learn programming or build a complicated AI system.

Tools such as ChatGPT Work and Claude Cowork make it possible to experiment with this approach using natural language.

The important thing isn't the technology itself.

It's understanding how to give AI a job, the right context, useful instructions and sensible boundaries.

First: Don't try to build a "super agent"

When people hear the term AI agent, they sometimes imagine one enormous AI system connected to everything they own.

That's probably not the best place to start.

Instead, think about the work you already do.

Maybe you spend too much time:

• Sorting emails

• Preparing meetings

• Writing reports

• Researching information

• Creating presentations

• Organising notes

• Analysing documents

• Preparing for interviews

• Following up with customers

Pick one.

That's your first AI agent.

The basic idea

Whether you're using ChatGPT Work or Claude Cowork, the principle is similar.

You give the AI five things:

1. A job

What do you want it to do?

2. Context

What information does it need?

3. Tools

What files, applications or sources can it work with?

4. Output

What should the finished result look like?

5. Boundaries

What is it not allowed to do?

This is essentially the prompting skeleton behind the 4 C's approach.

And it works for almost any AI workflow.

Using ChatGPT Work

ChatGPT can be used as a workspace for building repeatable workflows around your work.

Rather than opening a completely new conversation every time and explaining the same thing, you can establish the role, instructions and context for the work you're doing.

For example, imagine you want to create a Coordination Assistant.

Instead of simply saying:

"Help me manage my emails."

Give it a much clearer job.

You could start with something like:

You are my work coordination assistant. Your job is to help me organise incoming information, identify priorities, prepare responses and help me plan my working day.

When reviewing information, identify urgent tasks, important but non-urgent information and items that don't require my attention.

When you identify an email requiring a response, draft a response but do not send it.

When reviewing my schedule, identify potential conflicts and meetings that require preparation.

Present your recommendations clearly and concisely.

Never make commitments, send communications or change appointments without my approval.

Notice what we're doing.

We're not programming.

We're describing the job.

Give ChatGPT the context it needs

An AI assistant is only as useful as the context you provide.

If you tell it:

"Prepare me for my meeting."

it doesn't necessarily know what you need.

Give it more information.

For example:

• The meeting purpose

• Previous correspondence

• Relevant documents

• Customer information

• Previous meeting notes

• Your objectives

• Known problems

• Questions you want answered

Now the AI has something to work with.

Instead of:

"Prepare me for the meeting."

you can say:

"Review the previous correspondence and meeting notes. Identify the customer's main concerns, unresolved issues, decisions we need to make and anything I should prepare before the meeting."

That's a much better instruction.

The same principle works with Claude Cowork

Claude Cowork is particularly interesting for this approach because it is designed around working with information and files on your computer rather than simply having a conversation in a chat window.

That opens up a different way of thinking about AI.

Instead of:

Question → Answer

you can think:

Task → Files → AI works through the task → Result

For example, imagine you've received a folder containing:

• Several Word documents

• PDFs

• Excel files

• Meeting notes

• Customer information

You could give Claude a task such as:

"Review these files and create a summary of the project. Identify the current status, outstanding actions, important risks and any conflicting information. Create a concise management briefing."

That's much closer to the way you'd delegate work to another person.

Claude Cowork: Think in terms of a workspace

The big change in mindset is that you don't necessarily need to copy and paste everything into a chat.

Instead, think about the working environment.

You have a folder.

The folder contains information.

You give Claude a job.

Claude works with the available material and produces an output.

For example:

Project folder

📁 Project Alpha

→ Customer emails
→ Meeting notes
→ Contract
→ Budget spreadsheet
→ Project plan
→ Previous reports

Your instruction

"Review the contents of this project folder and prepare a management briefing. Highlight the current status, key risks, outstanding actions, upcoming deadlines and anything that requires a management decision."

That's a very different experience from asking:

"What is Project Alpha?"

The AI now has a job to perform.

Build the 4 C's in ChatGPT or Claude

Now let's bring the framework back.

You don't need four complicated systems.

You can start by creating four separate workflows.

C1 — Coordination

Your AI work organiser

Use this for:

• Emails

• Calendar information

• Priorities

• Deadlines

• Meeting preparation

• Follow-ups

Your basic instruction might be:

"Review my work information and identify what requires my attention today. Separate urgent tasks from informational items. Identify deadlines, meetings requiring preparation and potential conflicts. Create a prioritised action list."

Then add your boundaries:

"Do not send messages or make commitments without my approval."

The result is an AI assistant that helps you organise your day.

C2 — Creativity

Your AI production assistant

This is where you give AI raw material and ask it to turn that material into something useful.

For example:

"Here are my meeting notes. Turn them into a professional follow-up email, an action list and a short management summary."

Or:

"Turn this research into a presentation structure."

Or:

"Turn these notes into a first draft of a blog article."

The key is not asking AI to magically invent everything.

Give it your raw material.

Then let it do the heavy lifting of structuring and formatting it.

C3 — Clarity

Your AI research assistant

This is particularly useful when you have too much information.

Give the AI the documents and ask it to help you understand them.

For example:

"Review these documents and explain what I actually need to know. Identify the key decisions, risks, contradictions and unanswered questions."

For a single complicated document:

"Analyse this document section by section. Explain each section in plain English and identify anything that may require further clarification."

The important thing here is to distinguish between understanding something and making a final professional decision about it.

AI can help you navigate complexity.

For high-stakes legal, financial or professional decisions, you should still verify the conclusions and obtain appropriate expert advice.

C4 — Coaching

Your AI practice partner

This one requires very little setup.

Give the AI your context.

For example:

• Job description

• CV

• Company information

• Presentation

• Sales proposal

• Negotiation objectives

Then tell it who to be.

"Act as a skeptical hiring manager."

or:

"Act as a difficult customer."

or:

"Act as a CEO reviewing this proposal."

Then start the conversation.

The AI can challenge you with questions you might not expect.

Afterwards, ask:

"Review my performance."

Then:

"Where did I ramble?"

"Which answers were weak?"

"What should I have said?"

"Give me a stronger version of my answer."

Suddenly, you have a personal practice environment.

One prompt can become an entire workflow

Here's where things get really interesting.

Instead of asking AI to do one thing, you can describe the complete workflow.

For example:

I have a customer meeting tomorrow.

Review the relevant documents and previous correspondence.

Identify the customer's main concerns.

Summarise unresolved issues.

Identify information I need to have available during the meeting.

Create five questions I should ask.

Create five difficult questions the customer might ask me.

Prepare suggested responses.

Finally, create a one-page meeting briefing.

Do not invent facts. Clearly identify anything you cannot verify.

That's no longer just a question.

It's a workflow.

The secret is iteration

Your first prompt probably won't be perfect.

That's completely fine.

Start simple.

Then look at the result.

If it's too long:

"Make it shorter."

If it's too generic:

"Use the information from the documents rather than giving generic advice."

If it's missing something:

"Add deadlines and owners."

If the tone isn't right:

"Make it more suitable for a senior management audience."

You're effectively training the workflow through conversation.

Over time, you build instructions that consistently produce better results.

Don't automate the final decision

This is perhaps the most important rule in the entire approach.

AI can:

Find → Analyse → Organise → Draft → Recommend

You can:

Review → Decide → Approve

For example:

Email

AI drafts it.

You send it.

Calendar

AI identifies a conflict.

You decide what to change.

Contract

AI highlights a potential issue.

You decide whether to ask a lawyer.

Business decision

AI identifies the pros and cons.

You make the decision.

This keeps the human in the loop where it matters.

Start with one real task

Don't spend three days designing the perfect AI agent.

Pick something you do every week.

For example:

Every Monday I spend 45 minutes preparing my weekly priorities.

Give the AI the information it needs.

Ask it to create the first version.

Then improve the workflow.

Next week, do it again.

Once you have something that works, move to another task.

Eventually you might have:

My AI work system

Coordination

→ Organises my day

Creativity

→ Creates first drafts

Clarity

→ Helps me understand complex information

Coaching

→ Helps me prepare and practise

And suddenly you're not just "using ChatGPT."

You're building a personal AI-powered working system.

ChatGPT or Claude?

You don't necessarily have to choose one.

Different AI tools have different strengths, interfaces and workflows, and those capabilities continue to evolve.

The more important question is:

What job am I trying to get done?

If you're already comfortable with ChatGPT, start there.

If your workflow involves working extensively with files and your computer, Claude Cowork may be particularly interesting.

And there's nothing wrong with using both.

Your goal isn't to become loyal to a particular AI platform.

Your goal is to build a better way of working.

You don't need to be technical

This is perhaps the most exciting part.

The old way of automating work often required:

Software → APIs → Integrations → Code → Testing → Maintenance

AI is making another approach possible:

Describe the job → Give AI the context → Define the output → Set the boundaries

That's a huge change.

You can start experimenting with AI agents simply by learning how to delegate clearly.

And perhaps that's the real skill we need to develop in the AI era.

Not learning how to do everything ourselves.

Learning how to tell an AI assistant what good work looks like.

Your first AI agent

Don't build four.

Build one.

Find the most repetitive part of your working day.

Then write down:

JOB: What do I want AI to do?

CONTEXT: What does it need to know?

TOOLS: What information or files can it use?

OUTPUT: What should the finished result look like?

BOUNDARY: What must it never do without my approval?

Give those five things to ChatGPT Work or Claude Cowork.

Try it.

Improve it.

And then use it again.

That's how you move from using AI occasionally to actually working with AI.

And that's where the productivity gains start becoming much more interesting.

How to Actually Set Up Your First AI Agent

Let's make this practical.

You don't need to understand programming or APIs to start experimenting.

The easiest way is to create a dedicated workspace for a specific job and give the AI clear instructions.

I'll use a Coordination Agent as the example because it's one of the easiest and most useful places to start.

Option 1: Build it with ChatGPT

The exact names and available connections can change as ChatGPT evolves, but the basic process is straightforward.

Step 1 — Create a dedicated Project

Open ChatGPT and create a new Project.

Give it a name that describes the job rather than simply calling it "AI".

For example:

My AI Work Coordinator

This is important because you want the project to have a clear purpose.

You could eventually have separate projects such as:

• AI Work Coordinator

• Content & Creativity

• Research & Clarity

• Interview & Coaching

Think of each Project as a dedicated workspace for a particular type of work.

Step 2 — Give the Project instructions

Open the Project's instructions and tell ChatGPT what its job is.

Don't overcomplicate this.

Start with something like:

You are my personal work coordination assistant.

Help me organise my working day, identify important tasks, prioritise information, prepare emails and help me prepare for meetings.

When reviewing information, classify items as:

Urgent — requires my attention
Important — needs attention but isn't urgent
Informational — useful to know but doesn't require action
Ignore — doesn't require my attention

When you identify an email that requires a response, draft a response for me.

Help identify deadlines, meetings requiring preparation and potential scheduling conflicts.

Keep your recommendations concise and practical.

Never send emails, make commitments, cancel meetings or change appointments without my explicit approval.

That's already enough to create a useful starting point.

You can improve the instructions later.

Step 3 — Add your context

Now think about what your AI assistant needs to know about you.

For example:

• Your role

• Your responsibilities

• Your working hours

• Your priorities

• Important projects

• How you like emails written

• Your preferred tone

• Things you consider urgent

• Things you don't want AI to spend time on

You could tell it:

"I normally work from 09:00 to 18:00. I prefer concise emails and I want customer issues prioritised over internal administrative tasks."

The more relevant context you provide, the more useful the recommendations become.

Step 4 — Connect your work information

This is where the workflow becomes much more powerful.

Depending on your ChatGPT plan, workspace and the integrations currently available to you, you may be able to connect services such as Gmail, Google Calendar, Outlook or Microsoft 365.

The principle is simple:

Give the AI access to the information it needs to perform the job.

For a Coordination Agent, email and calendar are particularly useful.

Your email tells it:

What needs to happen.

Your calendar tells it:

When you have time to do it.

Together, they provide much more useful context.

Step 5 — Try a simple command

Don't immediately ask the AI to run your entire life.

Start with something simple.

For example:

Review my emails and identify anything that requires my attention today. Categorise each item as Urgent, Important, Informational or Ignore. For anything requiring a response, draft a suggested reply but do not send anything.

Now look at the result.

Ask yourself:

Did it understand what I consider important?

If not, change the instructions.

This is how you improve the agent.

Step 6 — Bring in your calendar

Once the email workflow is working, add your calendar into the process.

Now you can ask:

Review today's calendar together with my important emails. Identify meetings that require preparation, deadlines I need to be aware of and any potential conflicts between my workload and available time. Create a prioritised plan for today.

This is where the agent starts becoming genuinely useful.

It isn't simply reading your email.

It is connecting information.

Step 7 — Ask it to prepare your morning briefing

Once you've tested the basic workflow, turn it into a repeatable routine.

For example:

Create my morning work briefing.

Include:

1. My three most important priorities today.

2. Urgent emails.

3. Meetings I need to prepare for.

4. Important deadlines.

5. Potential calendar conflicts.

6. Any tasks that could become problems if I don't deal with them today.

Keep it concise and tell me what I should focus on first.

Now you've created something that can become part of your daily routine.

What about Claude Cowork?

Claude Cowork follows a slightly different approach.

The big advantage is that you can think of Cowork as an AI assistant working with the files and information in a designated workspace on your computer.

Instead of simply having a conversation, you can give Claude a job to complete using the information available to it.

Step 1 — Open Claude and enter Cowork

Open Claude and select Cowork.

The interface may change over time, but the concept is simple:

You're switching from ordinary conversation to a workspace where Claude can work on tasks and files.

This is where you can start thinking of Claude less like a chatbot and more like a digital colleague.

Step 2 — Create or choose a dedicated workspace

Create a folder on your computer for the work you want Claude to handle.

For example:

AI Work Coordinator

Inside it you might eventually have:

📁 Emails
📁 Meeting Notes
📁 Projects
📁 Reports
📁 Reference Documents

You don't need to build an elaborate folder structure.

Start with whatever makes sense for your work.

Step 3 — Give Claude a clear job

Now give Claude the same type of instructions you gave ChatGPT.

For example:

You are my work coordination assistant.

Review the information I provide and help me identify priorities, deadlines, outstanding actions and items requiring my attention.

Organise information into Urgent, Important, Informational and Ignore.

Draft responses when appropriate.

Never send communications or make commitments without my approval.

Do not invent information. If something is unclear, tell me.

Again, notice how simple this is.

You're not programming Claude.

You're delegating a job.

Step 4 — Give Claude access to the relevant files

Now put the documents it needs into the appropriate workspace.

For example, you could provide:

• Meeting notes

• Reports

• Project documents

• Spreadsheets

• PDFs

• Draft emails

• Research

Then ask Claude to work through them.

For example:

Review the project documents in this folder and create a management briefing showing the current status, outstanding actions, deadlines, risks and decisions that need to be made.

Claude can then work through the available material and create the output.

What about Gmail or Outlook?

This is where it's important not to confuse working with local files and connecting directly to an email account.

If you want your AI assistant to work directly with your Gmail or Outlook inbox, use the email integration/connectors that are actually available in the AI platform and your account.

Don't simply give an AI your email password.

Instead, use the platform's official connection process, where available, and review exactly what permissions you're granting.

For a first experiment, you can also keep things simple:

Export or copy the relevant information into your workspace and let the AI work on that.

Once you're comfortable with the workflow, you can explore deeper integrations.

The same 5-part structure works in both

Whether you're using ChatGPT or Claude, think about your agent like this:

JOB

"Manage my daily work priorities."

CONTEXT

"Here is my role, my priorities and my working style."

INFORMATION

"Here are my emails, documents and calendar information."

OUTPUT

"Give me a concise daily briefing with priorities, deadlines and preparation requirements."

BOUNDARY

"Don't send anything or make commitments without my approval."

That's your AI agent.

It really can be that simple.

Now build the other three C's

Once you've built your first Coordination Agent, the same approach can be used for the other three.

Creativity Agent

Create a project called:

AI Content & Creativity

Give it instructions such as:

"Turn my rough notes into professional first drafts while maintaining my writing style. Ask for clarification when necessary and never invent facts."

Then give it:

• Notes

• Previous articles

• Presentations

• Reports

• Brand guidelines

And ask it to produce the first draft.

Clarity Agent

Create:

AI Research & Clarity

Give it instructions such as:

"Help me understand complex information. Summarise documents, identify important issues, compare information and clearly distinguish facts from assumptions."

Then give it your:

• Reports

• Contracts

• Research

• Technical documents

• Business information

And ask it to create a structured analysis.

Coaching Agent

Create:

AI Coaching

Give it instructions such as:

"Act as a challenging but constructive coach. Ask realistic questions, challenge weak answers and provide specific feedback after each practice session."

Then give it:

• Your CV

• Job description

• Presentation

• Sales proposal

• Interview information

And start practising.

Don't try to perfect it on day one

This is probably the most important advice.

Your first AI agent will probably be mediocre.

That's okay.

The process is:

Build → Test → Correct → Improve → Repeat

Maybe the AI is prioritising the wrong emails.

Tell it.

Maybe the summaries are too long.

Tell it.

Maybe the writing doesn't sound like you.

Give it examples.

Maybe it keeps making assumptions.

Add a rule:

"If information isn't available, say that it isn't available. Do not guess."

Every improvement makes the system better.

Start with one agent

You don't need four AI agents tomorrow morning.

Start with one.

I'd recommend the Coordination Agent because it addresses one of the biggest problems in modern work:

Too much information competing for your attention.

Get that working.

Then build the Creativity Agent.

Then Clarity.

Then Coaching.

Eventually you have something much more valuable than a collection of clever prompts.

You have a personal AI working system.

And remember: you're still the boss

The objective isn't to hand your working life over to AI.

It's to delegate the repetitive parts.

AI can:

Organise.

Research.

Summarise.

Draft.

Analyse.

Prepare.

Challenge.

You:

Review.

Decide.

Approve.

Take responsibility.

That's the balance that makes AI useful.

You don't need to become a programmer.

You don't need to understand every technical detail.

Start with a task you already do every week.

Give the AI a clear job.

Give it the information it needs.

Tell it what the result should look like.

Set boundaries.

Then see how much time you get back.

That's the real beginning of working with AI.

Contact

Questions? Reach out anytime.

Email

automation@trabalhoia.com

© 2025. All rights reserved.