AI Skills Assessment Practice: Two Tasks and a Work Log

These tasks practice producing an accurate deliverable and checking the AI draft that led to it. Record the errors you found, the evidence you used and what you corrected, so you can explain both the final answer and your process.

One educational framework you can borrow vocabulary from is Anthropic's AI Fluency Framework, which separates Delegation (deciding what to hand off), Description (communicating the request), Discernment (judging the output) and Diligence (using the tool responsibly) official overview. The tasks below use those ideas loosely. They are Preterview's own practice material, not an official rubric from any employer.

Real assessments differ in allowed tools, time limits and submission format. Read the instructions you were sent first, then adapt these tasks to that format rather than practicing a format nobody asked for.

The six messages below are a fictional example created for this guide, from an imaginary online kitchenware shop. The rules: four categories (Shipping, Refund/Exchange, Account/Login, Product question) plus an urgency flag. Urgent is defined as 'a duplicate charge occurred, or the customer cannot use the service at all'.

The deliverable is one table with a category, an urgency flag and a one-line reason per inquiry. Where the rules do not settle the case, mark it 'needs a decision' and say why instead of guessing. Getting a draft from an AI tool and then verifying it is the whole task.

When you brief the tool, paste the rule definitions and the six messages verbatim, specify the output format, and add one constraint: do not add facts that are not in the message. That constraint gives you something concrete to check for later.

Verify one row at a time against the rule text. As a fictional example for illustration, suppose the tool returned the following draft.

Separate invented evidence (row 1) from incorrect application of the category or urgency rules (rows 2, 4 and 6). Keep the AI draft, correction and reason side by side. Checking whether the stated rule fits the business is a separate task from applying it correctly.

Row 6 reports neither a duplicate charge nor inability to use the service, so the stated rule makes it not urgent. A separate business question remains: should an address change receive priority before shipment? Record that proposed rule change separately from the current classification.

The second task checks whether a summary matches the numbers it claims to describe. The table below is a fictional example of three months of inquiry counts.

Suppose an AI tool produced this summary from the table: 'Total inquiries over three months were 720. Shipping inquiries rose 50% from June to August. Refund/Exchange inquiries increased every month. The drop in account inquiries in August is the result of the login page redesign in July.'

The deliverable is a list that tests each claim against the table, plus a rewritten summary that states only what the table supports. For each claim, record the claim, the value you computed and a verdict of match, mismatch or unsupported.

1. Total of 720: the monthly totals are 195, 225 and 272, which sum to 692. Mismatch.
2. Shipping up 50%: in the fictional table, June has 120 inquiries and August has 180. (180−120)÷120×100 = 50%, so the claim matches.
3. Refund/Exchange up every month: 45, 40, 60 includes a drop in July. Mismatch.
4. Account drop caused by the redesign: 35 to 32 is a drop, but the cause is not in the table. Unsupported.

The rewritten summary should stop where the data stops, for example 'Account inquiries fell from 35 in July to 32 in August'. If a cause matters, write a separate note saying what additional data you would need to support it.

A work log helps you explain how you produced and checked the deliverable. This is Preterview’s practice template; use the required submission format in an actual assessment.

• Task and goal: one sentence on what had to be produced
• Tool used and why: tool name and why it fit this task
• What I asked the AI for: rules, input data, output format, constraints
• Accepted as-is: which parts you kept and how you confirmed them
• Corrected: original value, corrected value, reason
• Discarded: what you removed and why
• Not verified: what remains uncertain and why
• Questions about the rules: where the rules were silent or ambiguous

As a fictional example, the 'Corrected' field for Task 2 might read 'Changed total from 720 to 692. Confirmed by adding the monthly totals myself.' A value paired with a method can be read aloud as an answer without reconstructing anything from memory.

If there is a debrief after the task, expect questions about your judgment rather than your result. These four are worth preparing.

• How did you split the work between the AI tool and yourself?
• Did you discard any of the AI output? What and why?
• Was there anything you could not verify? How would you have checked it with more time?
• How did you handle places where the rules were unclear?

A fictional example answer to the second question: 'The reason for inquiry 1 said the delay was due to low stock, but the message only said the order had not shipped, so I removed that. I had told the tool not to add facts, and it did anyway, so I re-read the other five rows for the same problem.'

Likely follow-ups are 'Did the same thing happen elsewhere?' and 'How would you change the brief to prevent it?' If your log has the Discarded and What-I-asked-for fields filled in, both questions can be answered from the page in front of you.

Run the deliverable and the log through this list when you are done. It is Preterview's suggestion for practice and may not match how any particular assessment is graded.

• Does the deliverable use the requested format (table, list, summary)?
• Did you find each type of error in the AI draft and record why you changed it?
• Is anything left in the deliverable that the source message or table does not support?
• Did you leave gaps in the rules as written questions instead of quiet decisions?
• Did you record what you could not verify rather than omit it?

If you are on the hiring side and want to run tasks like these, design the shared questions and the way evidence is recorded before the first candidate arrives. That is covered in Structured interview scorecard template.

If you are a candidate, the remaining step is saying your reasoning out loud. Preterview's document-based mock interviews and post-interview feedback can be used for that rehearsal. See pricing and terms and a sample report for what is currently offered.