Home › Blog › Build an AI agent with Claude

How to build an AI agent with Claude, without coding

Akash Gupta Founder, AI Gurukul Last reviewed 23 October 2026 10 minute read

I do not write code. I studied at IIM Ahmedabad and built a YouTube channel, not software. Yet today my company runs on AI Employees that I built in Claude Code, and each one is checked by a person every day. The skill that made this possible was not coding. It was learning to describe a job so clearly that a machine could do it.

This guide shows that skill with one small example you can build today. If you want the bigger picture first, read what an AI Employee is.

What is an AI agent, in plain words?

An AI agent is an AI that does a job for you, not just answers a question. A chatbot waits for you to ask, replies once, and stops. An agent has a job description, knows where to find what it needs, uses tools to do the work, and hands you a finished result to check.

There is a longer comparison in AI agent or chatbot: what is the difference?

Which Claude product should I use to build an agent?

Claude comes in a few forms, and people mix them up. Here is what each one is good for, checked against Anthropic's own help pages and docs on 23 October 2026:

• The Claude app, with Projects. A Project is a workspace inside the Claude app where you save instructions and upload files. Every chat in that Project follows those instructions and can read those files. Anthropic says Projects are available to all users, including free accounts, and free users can make up to five. One of its own pages also says project instructions are for paid plans only, so if the instructions box is missing on a free account, that is why. This is the easiest place to start. Source: Claude Help Center.
• Connectors. A connector links Claude to another app, such as Gmail, Google Drive or Google Calendar, so Claude can read from it and, if you allow it, act in it. Anthropic lists web connectors as available to all users. On Team and Enterprise plans, an owner must switch a connector on for the company first. Source: Claude Help Center.
• Claude Code. This is Claude working on a folder on your computer. It can read and write files, follow a saved instructions file, run saved routines called skills, and use the connectors you added in the Claude app. It runs in a terminal window, or inside the Claude desktop app under the Code tab, with no terminal at all. Anthropic's docs say it needs a Pro, Max, Team or Enterprise plan, and the free plan does not include it. Source: Claude Code docs.

My advice: build your first agent in a Project, so you learn the six parts with nothing to install. Move it to Claude Code when you want it to work on your own files and repeat the job on one short command. There is a full beginner guide in Claude Code for people who do not code.

What are the six parts of an AI agent?

In AI Gurukul, I teach every agent through the same six parts. I call it the AI Employee Blueprint. It works for any job, not just the example below.

How do I build an AI agent with Claude, step by step?

Let us build one small, useful agent: a Meeting Notes Helper. You give it the rough notes from a meeting. It gives back a clean list of who will do what, by when, and anything nobody owned. It needs no connector and cannot harm anything.

Part 1. Write the purpose

Open a notes app and write four lines. This is your agent's job description.

1. Job: turn rough meeting notes into a list of actions.
2. Goes in: notes I paste in, in any format.
3. Comes out: a table with the task, the person who owns it, the due date, and a line for anything with no owner.
4. Stops at: it does not send anything to anyone. I read the list and share it myself.

Part 2. Give it a brain

1. In the Claude app, open Projects and create a new project. Call it "Meeting Notes Helper".
2. Find the project's instructions box and paste your purpose from Part 1.
3. Under it, add the steps in order. For example: "Read all the notes first. List every action. For each one, name the owner only if the notes say who. Never guess a name or a date. If something has no owner, put it under Needs an owner. End with any question you could not answer from the notes."
4. Pick a model from the model menu in the chat. For a job like this, the standard model the app suggests is fine. You can try a larger one later if the results miss things.

Part 3. Give it memory

1. Upload one short file to the project's knowledge: the names and roles of the people who usually attend your meetings.
2. Upload one example of a "good" action list, made by you. Claude copies the shape of a good example better than it follows a long description.

Part 4. Choose its tools

For this agent, the honest answer is: none. Pasting notes in is enough. This is the most skipped lesson in agent building. Only add a tool when the job truly needs it. Later, you could connect Google Drive so it reads notes from a folder, but not on day one.

Part 5. Set the guardrails

Add three lines to the instructions:

1. "Never invent a name, a date or a number that is not in the notes."
2. "If the notes say something private, such as a salary, leave it out of the list."
3. "Do not write any message to send. Only the list."

Part 6. Test it

1. Paste the notes from a real meeting you attended, so you know the right answer.
2. Check every line against your memory of the meeting. Did it miss an action? Did it invent an owner?
3. Now try a messy case: notes with no names at all, or two people with the same first name.
4. Each time it gets something wrong, add one line to the instructions to fix it, and test again.

After three or four rounds, it does the job the same way every time. That is when a chat becomes an employee.

How do I move the same agent into Claude Code?

When you want the agent to read notes from a folder on your computer and save the list as a file, Claude Code is the next step. The six parts stay the same. Only where you write them changes:

• Brain: the instructions go in a plain text file called CLAUDE.md in your folder. Claude Code reads it at the start of every session. Anthropic's docs describe it as the place for "facts and standing guidance that Claude always loads".
• A repeatable routine: the steps can become a skill, which is a folder with a file called SKILL.md inside .claude/skills/. You then run the whole job by typing / and the skill's name. Source: Claude Code docs on skills.
• Memory: the files sit in the same folder, where Claude Code can read them.
• Tools: connectors you added in the Claude app show up in Claude Code by themselves when you log in with your Claude account. You can see them by typing /mcp. Source: Claude Code docs.
• Guardrails: Claude Code has permission settings that block an action outright. This matters, because a rule written in your instructions shapes what Claude tries to do, but the docs are clear that it is the permission settings, not the instructions, that decide what Claude Code allows.

You do not write these files by hand. Ask Claude Code to create them, then read what it wrote before you say yes.

How do I keep my AI agent safe?

Start every new agent so that it can only read, never change or send. Give it more power only when it has been right for weeks, and only one power at a time. In practice:

• Read only first. An agent that can only read can make a wrong list. It cannot send a wrong email.
• A person says yes. Anything that leaves your company, moves money or deletes something waits for you. Claude's Gmail connector, for example, asks for your approval before each send, reply or forward by default.
• Least access. Connect only the app the job needs. A notes helper does not need your email.
• Private data. Keep passwords, ID numbers and bank details out.

The safest first build with a real app is an email helper that only reads. I wrote a full guide on how to build an AI email assistant for Gmail that only reads.

What are the honest limits of a Claude agent?

• It can be wrong and sound sure. Claude can miss a line or misread a date. This is why testing and a human check are parts of the build, not extras.
• It only knows what you give it. A decision made on a phone call and never written down is invisible to it.
• It does not run by itself by default. In a Project, it works when you open a chat and ask. Running a job on a timer is possible in some Claude products, but it is a later step, and not one I suggest for a first agent.
• Plans and features change. Claude adds and moves features often. Everything in this guide was checked on 23 October 2026. Check the help pages linked above before you rely on a detail.
• It costs money for the serious version. Projects exist on the free plan, but Claude Code needs a paid plan.

For a wider view of which office jobs suit an agent and which do not, read how office workers use AI agents.

Akash Gupta, Founder, AI Gurukul

Akash studied at IIM Ahmedabad and built Peepoye, a Hindi YouTube channel. He does not write code. He runs his own company, Peepoye Network, with AI Employees he built himself with Claude, and teaches the same method in AI Gurukul. His story.

28.2 croreviews on his own channel

12.8 lakhsubscribers

1.4 crorehours watched