Ask for One Inspectable Action

You now have enough vocabulary to guide a small action without writing its technical instructions. Start with a familiar outcome: explain the first sentence of the prepared welcome note. The task has a single source, a small result and no reason to contact an external service. This is a useful model for gradually building confidence with an agent's tools.

The prepared small-task.md includes an ordinary-language request: “Read welcome-note.md in this course folder. Explain what a new customer learns from the opening sentence. Leave the file unchanged.” Notice the three ingredients. The source is named, the question is understandable and the boundary is visible. The assistant can choose a suitable reading tool without you specifying a command.

In a live session, the assistant might first identify the file, read it and then answer. Review the activity if your app displays it. Did it use the expected folder? Did it return information that matches the file? Did the final answer address the question? The exercise does not become better merely because the agent used more tools. A small, clear result is the point.

If the assistant asks for access that does not seem connected to this simple task, ask why it is needed. There may be an ordinary explanation, such as the file being outside the current working folder, but you should understand it before expanding the task. The course example deliberately stays inside prepared local material so you can focus on the working pattern.

After an inspection works, you might later choose a one-sentence draft or edit. That is a separate next step, not an automatic extension of the original read request. You can ask to see proposed wording in the conversation before deciding whether to change the file. This keeps each result small enough to assess while you learn what your chosen assistant can actually do.

The browser activity compares a vague request with the prepared one. You can complete it without an AI account. Optional app practice uses the same fictional file and does not require a terminal. By the end, you should be able to describe a small job in familiar language, recognize the tool action that supports it and check whether the result answered the request you actually made.