Your first AI CLI session should not begin with “build my startup.” It should begin with orientation: open one safe project folder, start one agent, and learn what it sees before asking it to act.

The terminal is not a test of whether you are technical enough. It is the place where an AI agent can see the files, commands, and outputs that make up a real project.

Before you start

Choose a folder that is safe to inspect. Do not begin in a folder full of private credentials, irreplaceable documents, or a production system you do not understand.

Check these four things:

The first request

Start the agent from the project root. Then give it this kind of request:

This is a better first session than immediately requesting a feature because it gives you a baseline. You can compare the agent's explanation with the actual folder. If it misunderstands something, correct the project documentation now.

Open project rootAsk agent to read and orientCompare explanation with real filesApprove one small planVerify result

Understand permission before convenience

AI CLI tools may ask permission to read files, edit files, run a command, install a package, or connect to another service. Read the request before approving it.

Do not turn off all approval prompts just to make the agent faster. Early prompts teach you what the agent is really doing.

The rhythm of one good session

1. Orient. Read context and map the project. 2. Plan. Ask the agent to state the files, approach, and checks. 3. Approve. Confirm the plan or tighten the scope. 4. Build. Let it make the focused change. 5. Verify. Run checks and use the browser yourself. 6. Record. Review the Git diff and update the handoff.

The agent may perform several commands inside the build step. That is fine. You should still be able to explain the larger loop afterward.

A safe first task

Ask the agent to create or improve a README.md. It can inspect the project and explain its purpose, install command, run command, test command, and folder structure. This is useful work, low risk, and immediately improves the next session.

Practical exercise

Open a project folder and run one orientation-only session. Save the agent's project map in a Markdown file. Then ask it to propose—not execute—the smallest improvement it could make. Practice saying “show me the plan first.”

Check your understanding

1. Why begin in the project root? 2. What should an agent do before changing an unfamiliar project? 3. Which actions deserve an explicit approval pause? 4. What is a low-risk first task for an AI CLI?