Building an AI agent with Claude means: create a project, write instructions, upload knowledge, and test on a real case. It all happens inside the Claude app, without a single line of code, and you're holding the first usable result after about an hour. Here's the process, exactly as I use it myself.
My entire workforce runs on Claude: eight AI employees, each with its own job. What you'll build here is the basic form of that. I've collected all the agent fundamentals on the AI agents page, and why I use Claude for this work instead of ChatGPT is in Claude instead of ChatGPT.
What you need
A Claude account is enough. Projects are included in the free plan too; unlimited projects only come with the Pro subscription at 20 dollars a month (as of August 2026). For the start, meaning one agent and real test cases, the free plan will do; its message limits get tight, though, as soon as you work with it daily.
More important than the subscription is your preparation: pick a task that repeats, whose result you can judge in minutes, and that noticeably costs you time. For this guide I'll use "preparing proposals" as the example; the process is the same for any task.
Step 1: Create a project
Create a new project in Claude and give it a name that describes the job, for example "Proposal Agent." A project is its own workspace with two properties that turn a chat into an agent: its own instructions, which apply to every conversation inside it, and a knowledge base that every conversation can access.
That solves the core problem of the regular chat window: you no longer have to copy the context together for every task. It lives in the project and stays there.
One project per job, that's the rule. Don't stuff "everything about my business" into one workspace. An agent with a clearly defined job delivers better results and is much easier to correct than a jack-of-all-trades.
Step 2: Write the instructions
In the project instructions, you write down what the agent should do and how. No prompt engineering, just what you'd tell a new hire on their first day:
- The job: "You prepare proposals. You get my conversation notes and deliver a finished proposal draft."
- The process: "First clarify services and scope, then prices according to my price list, then the text based on my examples."
- The boundaries: "You never invent prices or services. If you're missing a piece of information, you ask instead of guessing."
- The format: "Deliver a finished document, structured like my example proposals."
The last point under "boundaries" is the most important one. An agent that asks instead of guessing saves you the rework that makes so many AI results unusable.
Step 3: Upload knowledge
Now you fill the project's knowledge base with what the agent needs for the job:
- your price list or pricing logic
- two or three real proposals from you as style templates
- a short description of your business: services, target audience, typical cases
- phrases you would never use
You don't have to polish any of this. Existing documents are enough to start. A trick from my own practice: let the agent do the asking. In the first conversation, write "Ask me all the questions you need for your job," and answer them. Then put the answers into the knowledge base as a document. That way the knowledge comes out of your head without you staring at a blank page.
Step 4: Test on a real case
Don't use a made-up example; use the proposal that has to go out this week anyway. Give the agent the conversation notes and let it work.
The first result won't be perfect, and that's the point of the test. What matters is what you do with the mistakes: don't rewrite it yourself, name them. "The price on line item 2 is wrong," "too formal, look at example 3," "the opening is missing." The agent corrects, you compare.
Resist the reflex to quickly rescue the result yourself. Every correction you silently make on your own is one the agent will get wrong again on the next case, because it never heard about it.
Step 5: Make corrections permanent
This is where the agent parts ways with the one-off prompt. Every correction that should apply more than once goes back into the instructions or into a document in the knowledge base. "Too formal" becomes a line in the instructions: "You write casually, no bureaucratic language."
After three to five rounds, you'll notice the rework approaching zero, because the corrections no longer live in your head, they live in the project. From here on, every new case is routine: notes in, draft out, approve.
From agent to employee
What you have now is a working AI agent: an AI with a job, context, and a repeatable result. What it's still missing is the management layer, which in my view makes the difference: a name, a proper personnel file instead of scattered documents, a clean onboarding, and feedback as a fixed part of working together.
That's not a technical step, it's a shift in perspective: you're no longer operating a tool, you're managing an employee. The complete process from choosing the task to the permanent hire is in Hiring an AI employee: the process.
And in case you're wondering why I build agents in Claude instead of n8n or Make: the reasoning, with a comparison table, is in Building AI agents without code: my way. The short version: workflow platforms solve assembly-line tasks, but the expensive hours of solo business owners are in tasks that need context and judgment, and for those, the project approach is the more direct path.
Where to start
Take an hour, a real task, and the five steps above. If you're unsure about the money: what the free plan can do and when the Pro subscription pays off, I've written up honestly in Build an AI agent for free.
