Build an AI skills test for job applicants

What you need

• The job, and three tasks this person would perform in a normal week
• A Claude account (Free works, because Claude Docs is on every plan and candidates open the exported Word file without a Claude account)
• The setting "Cloud code execution and file creation" turned on. You'll find it in Settings, then Capabilities.
• An email address where candidates send their answers
• A current employee or a friend to take the test once

The test goes out as a Word file instead of a link because of how sharing works on the free plan. Older guides tell you to publish the test in Claude and send the link. The Publish button only appears on artifacts made in a chat before September 16, 2026. On Free, you can't share anything newer by link, and anyone who opens a shared one needs a Claude account.

A free artifact also can't save what someone types into it, since that takes a paid plan. A Word file avoids both problems. Candidates type their answers in the file and email it back to you.

The prompt

Replace the words in [brackets] with your own. Good tasks include summarizing a long vendor email thread, turning a messy list into a schedule and drafting a reply to an upset customer.

The planted mistake is the core of the test. Anyone can paste a task into an AI tool and paste back the output. What you're testing is whether they read the answer before they hand it to you.

Steps

1. Pick three tasks from the real job, the kind this person would handle in a normal week.
2. Paste the prompt and read the test Claude builds. Ask for changes until it sounds like your business, or click into the doc and type your own fixes.
3. Read the scoring guide and rewrite anything you wouldn't stand behind. Before the first candidate sees the test, decide what score moves someone to the next round. An average of 3 out of 4 counts as a B or better.
4. Try the test with a current employee and time it.
5. Export the test. Click "Export" on the doc and choose Word (.docx). Open the file and make sure the scoring guide isn't in it.
6. Send the same file to every candidate for the role, with the same instructions and the same deadline. Tell them AI helped you write the test and a person scores it. If a candidate asks for an accommodation, like extra time or another format, work with them.
7. Score each test yourself with the guide, then make the call. Don't paste anyone's answers into Claude to grade or rank them.

Check the result

• The test fits in 20 minutes for someone who knows the work.
• All the sample material is made up.
• The scoring guide grades the work and how they checked it, never personality or how someone comes across.
• Every candidate got the same test, the same instructions and the same deadline.
• The scoring guide isn't in the file candidates get, and the trial run's answers arrived in your email.
• You scored every test and decided who moves on. Claude didn't score or rank anyone.

What a finished test looks like

Here's a sample for a made-up job: office manager at an insurance agency with three offices. Every detail is invented. Use it to see the structure, then build yours from the real job.

Sample test, top of the page: "Office manager AI skills test. This takes about 20 minutes. Use any AI tool you like; that's expected. Under each task, put your final answer, then two or three lines on how you checked the AI's work. Save this file and email it back by Friday at 5 p.m. A person scores every test. If you need more time or another format, just ask."

Sample task 1: summarize a vendor thread. The material is three short emails. In the first, the office supply vendor says copy paper goes from $38 to $42 a case on Nov. 3. A week later, a second email says the new price "starts Nov. 10" and moves Thursday deliveries to 7 a.m. The agency buys six cases a week.

The task is a five-line summary for the owner covering what changed, what it costs and what needs a decision. The catch is the two start dates. A good answer flags them and calculates the cost correctly: $4 more a case times six cases is $24 a week.

Sample task 2: build a visit calendar. The material is a messy list of eight vendor visits and deliveries for the three offices. The task is to turn it into one weekly calendar and flag any two visits that land at the same office in the same hour. The catch is that the copier technician and the new 7 a.m. paper delivery both arrive downtown on Thursday morning.

Sample task 3: answer an upset client. A client writes that a portal outage on the due date left them with a $25 late fee. The policy note says the payment portal was down from 6 to 8 p.m. Those fees come off the next bill, and the office gives no cash refunds. The catch is that AI drafts tend to promise refunds. A good answer stays inside the policy.

Sample scoring guide, one row (task 1):

Look at what that row grades: the work, and how they checked what the AI produced. Nothing in it scores personality, confidence or how someone comes across.

If it goes wrong

• Claude answers in the chat and never makes a doc. Check that "Cloud code execution and file creation" is on in Settings, then Capabilities. Start your message with /docs, or click "Output" in the message box and choose "Docs."
• You can't find a "Publish" button. That's expected for anything made since September 16, 2026. Use "Export" instead.
• A candidate says the link asks them to log in. You sent a Claude share link instead of the file. Export the doc to Word and attach the file.
• The scoring guide is inside the test. Ask Claude to move it to its own doc, then export again. Open the file before you send it, every time.
• A candidate can't copy the sample material into their AI tool. A Word file has no copy button. Tell candidates to open the file on a computer, select the text and press Ctrl+C (Cmd+C on a Mac).
• A candidate used ChatGPT, Gemini or Copilot instead of Claude. That's fine. The point is checking the work, not the brand of the tool.
• The trial run took 35 minutes, not 20. Cut the material or drop to two tasks, then run it again.
• A candidate doesn't have an AI account. Claude has a free plan. Say so in your instructions.
• You hit the free plan's limit while building. It resets every five hours. Resume where you stopped in the same chat.
• You want to change the test after it's gone out. Leave it alone for this round so everyone gets the same test, and fix it before the next one.

Common questions

How do I test a job candidate's AI skills?

Give every candidate the same short task from the real job. Tell them to use AI, and ask them to show how they checked what it produced. Then score the answers yourself with a guide you wrote before anyone took the test.

What should an AI test look like for a job that isn't in tech?

It should look like the job. For an office manager, that means emails, schedules and customer replies. For a bookkeeper, it's a messy expense list. Nobody should need to write code to pass an office test.

How do I know if a candidate's AI answer is wrong?

Plant the mistake yourself, such as two dates that don't match or a number that doesn't add up. Then you know what a careful answer catches, and you aren't grading on a hunch.

Can candidates use ChatGPT instead of Claude?

Yes, they can. The test grades the work and how they checked it, not which tool made it. Ask them to name the tool in their notes so you know what you're looking at.

Ask for a portfolio, too. Some candidates will already have work to show: scripts, Claude skills, small programs and pieces they made with generative tools. Ask them to bring it to the interview or send it shortly after. The work that impresses me most is batch files and automations, shortcodes, widgets and apps built from real design files, and scripts in Excel, Google Sheets or anywhere else in Google Workspace or Microsoft 365.

Hiring rules vary by state and city. This post isn't legal advice; for anything specific to your business, check with your state labor department or an employment attorney.

The test shows how someone works with AI. To see whether they fit the job itself, use The job-fit questionnaire. It follows the same rules: the same questions for everyone, a scoring guide, and you make the call.