Frequently asked questions

Agentic AI refers to AI systems that work toward a goal over several steps on their own: they plan, use tools such as reading files or running programs, look at the result, and decide what to do next. Anthropic's engineers describe agents as systems in which the model directs its own process and tool use, as distinct from workflows whose steps are fixed in advance (Building effective agents, Anthropic).

A coding agent is agentic AI applied to software: you describe a tool in plain language, and the agent writes the files, runs them, checks whether they work and fixes what fails, asking your permission for actions that matter.

A chat assistant answers in a conversation; if it writes code, you copy it somewhere and run it yourself. A coding agent such as Claude Code works inside a folder on your computer: it creates and edits the files, runs them, reads the errors and tries again. The result is a set of ordinary files on your laptop, such as a web page or a dashboard. The coding agent basics page explains this in about 15 minutes.

No. The workshop is designed for people who have never written code. You describe what a tool should do, check what the agent builds, and ask for changes in plain language. No terminal is needed: the workshop uses the Claude desktop app, set up in the setup guide.

Anything you want: one small working tool from your own idea, built in about 40 minutes and shown in a 90-second presentation. It can be a web page, a form, a dashboard, a calculator, a chatbot, a checklist, a game or anything else the problem in your worksheet calls for. If your tool works with records, you test it with records you make up; the agent can create them.

Yes. Everything the agent makes is saved as files in the project folder on your laptop. If you build a website, you can also publish it with a free GitHub account after the workshop.

Claude Code is not included in the free Claude plan; it starts with the Pro plan (Claude pricing). A coding agent also does far more work per request than a chat reply: it reads files, writes code, runs it and checks the result, often many times for one request, and the paid plans cover that usage.

As of October 2026, Pro costs $20 a month when billed monthly. Prices change, so check the pricing page before you subscribe.

Not if you use U-M's Claude Code access for faculty and staff. Follow the sign-in instructions you received and skip step 1 of the setup guide. Students need a personal paid plan.

They work in a similar way, but the instructions and help on the day assume Claude Code. If you use another agent, you will need to adapt the steps yourself.

A Mac with macOS 13 or later, or a Windows laptop with Windows 10 (version 1809) or later. A Chromebook or a tablet cannot run Claude Code; if that is all you have, reply to the workshop email before the workshop.

Start with the troubleshooting table at the end of the setup guide. If that does not solve it, reply to the workshop email with a screenshot of the terminal, so the problem can be fixed before the workshop.

No. The Claude desktop app includes Claude Code: in its Code tab you open a folder, type requests and approve changes with buttons. The workshop uses the desktop app. People who prefer the terminal can use it instead; the setup guide has an optional section for it.

No. Installing the app and signing in does not upload your files. Claude works only with the folder you open in a Code session, and only what you type and the files it reads there are sent while you work, as described under "What does Claude Code send to Anthropic?" below. Files elsewhere on your laptop are not read.

Yes. You need it to keep your project online and publish it, and students, faculty and researchers can apply for GitHub Education benefits with it (staff in other roles may not qualify). The setup guide, step 5, covers both.

No. Use only records you make up; the agent can create realistic ones for testing. No client, field placement or employer data go into the agent. A prototype built in a few hours has not been checked for the safeguards real client data require.

Claude Code runs on your laptop, but to work it sends your requests and the model's responses, including the parts of files it reads, to Anthropic over an encrypted connection. On personal plans (Free, Pro and Max), you choose whether this data may be used to train future models; the setting is at claude.ai privacy settings. With the setting off, data are kept for 30 days; with it on, for five years. Details: Claude Code data usage.

Only in your browser. The worksheet page sends nothing anywhere; download the Word file to keep a copy.

The term was coined by Andrej Karpathy in February 2025 for building software by describing it to an AI and accepting the code it writes, with little or no reading of that code. It works well for quick experiments. In the workshop you write a short spec first and check what the agent builds against it, because a tool other people will rely on needs more than a good feeling about it.

Under Anthropic's consumer terms, you keep your rights in what you type in, and Anthropic assigns to you whatever rights it has in the output, "if any". Whether the output is protected by copyright at all is a separate question. The U.S. Copyright Office concluded in January 2025 that material generated purely by AI is not protected, and that prompts alone generally do not make you its author; the parts you write or substantially change yourself, such as your text, your spec and your edits, can be (Copyright and Artificial Intelligence, Part 2).

In practice, you are free to use, change and share what you build. For a specific question, such as a tool you plan to sell, U-M's Copyright Services can advise. This FAQ is not legal advice.

Borrowing ideas is common and acceptable: layouts, features and ways of organizing information are ideas that sites take from each other all the time. The text, photos, illustrations, logos and code of a site belong to their owners, so do not ask the agent to copy them, download them, or rebuild a site page for page. Describe what you like instead: "a page with a search bar at the top and results as cards below, like a library catalog" gives the agent what it needs without copying anyone's work.

Only with permission or an open license, and with credit where the license asks for it. Your own photos and writing, images with a Creative Commons license, and simple icons from open icon sets (this site uses Font Awesome) are the safest choices. Organization logos, including U-M's, have their own rules of use.

The models behind coding agents learned from large amounts of public code, so output can occasionally resemble existing code; for a small custom tool the chance of copying a substantial piece is low. If a passage looks like it came from a particular project, ask the agent where it comes from and rewrite it, or follow that project's license.

Yes, as long as you follow each library's license. Most common ones, such as MIT, BSD and Apache 2.0, allow free use if you keep the copyright and license notice. Others, such as the GPL, require that a program which includes them and is shared with others be released under the same license. Ask the agent: "List every library this project uses and its license," and keep the license files it adds.

Only if you want others to reuse it. Without a license, default copyright law applies and others may not copy, share or build on your code (GitHub, Licensing a repository). If you want people to reuse it freely, a short permissive license such as MIT is common; choosealicense.com, run by GitHub, compares the options in plain language. The agent can add the license file for you.

Worldwide, the amount is large and growing. The International Energy Agency estimates that data centres used about 1.5% of the world's electricity in 2024 (about 415 terawatt-hours) and projects that this will roughly double to about 945 terawatt-hours by 2030, with AI the main driver (Energy and AI, 2025). Data centres also use water for cooling, and new ones put pressure on local power grids and water supplies.

For a single request, the amount is small but not zero. Google reported in August 2025 that a median text prompt to its Gemini apps used about 0.24 watt-hours of electricity and 0.26 millilitres of water, about the energy of nine seconds of television (Measuring the environmental impact of AI inference). The figure excludes the energy used to train the model, and researchers have noted that Google did not say how long a median prompt is. Other models and providers differ, and most do not publish figures per request.

Yes. One request to a coding agent can involve many rounds of reading files, writing code, running it and checking the result, so it uses more computing than a single chat reply. A few habits keep the use down without making the work harder:

• Write a clear spec first, so the agent builds the right thing once instead of several times.
• Ask for one change at a time in the same session, rather than starting over.
• Close sessions you are not using, and do small edits, such as fixing a typo, yourself.

Whether a tool is worth its footprint is a fair question to raise in your organization as well, together with cost, privacy and who maintains it.

A published site needs two things: hosting, a computer that holds the files and is always on (think of it as land), and an address people type to reach it. A website or web tool can be published with GitHub Pages, which gives it an address such as yourname.github.io/project. You need a GitHub account. You can ask Claude Code to do it: "Publish this folder as a GitHub Pages site and tell me the address." It asks before each step. GitHub's guide Creating a GitHub Pages site covers the same steps by hand. Vercel, the optional step 5 in the setup guide, is an alternative that connects to the same GitHub repository.

If hosting is the land, a domain is your own street address in place of the github.io one. Buy the name (for example, yourproject.org) from a domain registrar, which charges a yearly fee that varies by name and registrar. Then connect it to your GitHub Pages site:

1. In your repository on GitHub, open Settings, then Pages, enter the domain under "Custom domain" and save.
2. At the registrar, add the DNS records GitHub lists (DNS is the internet's address book, and these entries point your address to GitHub): a CNAME record for www pointing to yourname.github.io, and for the bare domain, four A records with GitHub's addresses.
3. Wait for the change to take effect (up to 24 hours), then tick "Enforce HTTPS" in the same Pages settings.

Details: Managing a custom domain for your GitHub Pages site. Claude Code can explain each record as you go, but you sign in to the registrar yourself. Services such as Cloudflare can also manage a domain's DNS; none of this is needed for the workshop.

Treat what you build in the workshop as a prototype. Before it handles real client information or becomes part of someone's work, talk to the organization's IT or privacy contact: a tool that stores or sends client data needs a security review, a decision about where the data live, and someone responsible for keeping it running.

On claude.ai or the desktop app, open Settings, then Billing, and choose Cancel. The plan remains active until the end of the billing period; cancel at least 24 hours before the next billing date to avoid the next charge. If you subscribed through the iPhone or Android app, cancel through that app store. Details: Cancel your Pro or Max subscription.

Anthropic's free course Claude Code 101, the Claude Code best practices guide, and the official videos listed on the basics page.