Claude Code: Build a reusable weekly AI agent

Key insights

• Use Claude Code in the desktop app and select This computer to work in a local folder.
• Put lasting business details and rules in CLAUDE.md; save repeatable jobs as skills.
• Plan mode lets you review a proposed plan before Claude changes files.
• Separate tasks into work it can do, drafts for your approval and tasks you keep.
• Subagents can handle tasks in parallel, but each uses some of your plan's limit.
• The demonstrated email workflow saves drafts; the user reviews and sends them.

Claude Code lets you build a reusable weekly AI agent that works from a folder on your computer, follows saved instructions and checks its work. This approach is for people who want to automate recurring tasks but keep control of actions such as sending email or spending money. Teacher's Tech demonstrates the setup using a small coffee roaster's Monday workload as an example.

What it does

The agent in the example sorts weekly chores into three groups: work it can do, work it should draft for approval, and work that needs the owner. It reads business files, handles tasks with helpers called subagents, and has a separate reviewer check completed work. The example also creates a custom Roast and Reorder board and saves the routine as a skill that can be run again.

That structure distinguishes an agent from a chatbot that returns suggestions for a person to carry out. The agent is given a goal, plans the work, uses available tools, checks results and continues until it reaches an outcome. The user still sets the boundaries, approves sensitive actions and reviews the result.

Who it is for and what you need

The demonstration uses Claude Code in the Claude desktop app on a Mac or Windows computer. The transcript says a paid Claude plan, Pro or higher, is needed and that the presenter used Pro. Gmail is optional and is needed in this example only for email drafts.

The work folder is central to the setup: the rulebook, saved skill and output files live there. The transcript advises using the desktop app with “This computer” selected, rather than the browser version, because the browser version described in the demonstration runs from a GitHub repository and cannot see a folder on the computer. Availability by region and any broader licensing details are not stated.

How it works

1. Create a home for the agent. In the Claude desktop app, select the Code icon, choose “This computer” and pick a new or empty folder. The example uses one folder for the weekly agent so its files and instructions stay together.
2. Write the rulebook. Ask Claude to create a file with relevant business details and standing rules. In the example, the rules include saving outputs in a folder, keeping messages short, never sending email and asking before spending money. The file is read at the start of sessions in that folder. The presenter emphasizes that it is guidance, not a hard technical lock.
3. Connect tools and sort tasks by risk. In the app, go to Customize, then Connectors, and connect Gmail if email drafting is needed. Decide which tasks the agent may complete, which it may prepare for review and which remain with the user. The example leaves orders and phone calls with the owner, while emails are drafts only.
4. Use Plan mode to define the job. Switch from Auto to Plan before asking Claude to build the workflow. Give it the goal, why the task matters, the rules it must follow and what a finished result should include. Review the proposed plan, answer questions and make changes before accepting it.
5. Give specific feedback, then save the routine. If the first result misses the mark, identify what is wrong and what should change. Save a recurring workflow as a skill, such as , and run it with its slash command in later sessions. The rulebook holds standing instructions; the skill holds the steps for a particular job.
6. Have helpers work and a reviewer check. The main agent can assign individual chores to subagents that work at the same time. A separate reviewer checks each task against a clear checklist before it is marked complete. Subagents use some of the plan's limit, so the demonstration recommends starting with a few tasks rather than many.

For email, the example has two safeguards: the rulebook says drafts only, and Claude asks for permission before sending through Gmail. The owner opens Gmail Drafts, reviews the messages and sends them manually. A repeated run can start in a new session using the same folder and skill, so the routine and rules do not need to be re-entered.

Limitations and open questions

A rule in is not an absolute barrier. The demonstration distinguishes that guidance from permission prompts, which act as a hard stop for actions such as sending an email. Users should not treat the rulebook or a reviewer as a guarantee that every result is correct; the transcript shows the reviewer catching truncated coffee names and a broken link.

Connected services and source files must be available for the job to work as intended. The description notes that Gmail drafts may not appear if Gmail is not enabled for the session or contact records lack real email addresses; in that case, drafts may be saved as files. The local agent also needs the computer to stay on while it works. Phone access through remote control does not move the work to the phone, and the transcript says the session reconnects when the computer wakes.

The example does not establish availability across all plans, regions or organizations, and it does not provide a full comparison of Claude Code with Microsoft agent products. Treat the specific menu labels and plan details as those shown in the demonstration, not as a universal tenant or licensing statement.

Sources

• Teacher's Tech demonstration of building a Claude Code agent
• Teacher's Tech channel

Frequently asked questions

What do I need to build an agent in Claude Code?

The demonstration uses the Claude desktop app on Mac or Windows and says a paid Claude plan, Pro or higher, is needed. Gmail is optional if the agent needs to prepare email drafts.

Can Claude Code use a folder on my computer?

The demonstrated setup uses the Claude desktop app with “This computer” selected and a chosen folder. The transcript says the browser version cannot see folders on the computer.

How do you make a Claude Code workflow reusable?

Save the task's steps as a skill in the agent's folder, then run it with a command such as /brew-it. Keep standing instructions in CLAUDE.md.

Will a Claude Code agent send email by itself?

In the example, the rulebook says drafts only, and sending through Gmail requires permission. The user reviews and sends the email manually.