Knowledge Check: What It Is and How to Write One

Go from storyboard to an interactive course in minutes

Start Free

A new hire works through a 40-minute induction module, clicks Next on every page, then meets a 20-question quiz at the end. They pass with 85%, and a week later they cannot tell you what the module asked them to do on a live call. Nothing in the module asked them to remember anything while they were learning it.

A knowledge check is a short set of low-stakes questions placed inside or straight after a lesson, so learners recall what they just learned and get feedback before they move on. Retrieval can strengthen later recall. The feedback is what stops a wrong answer sticking too, so I judge a knowledge check by whether its wrong answers teach something, not by its score.

I'm an instructional designer at Coursebox, where I build and test courses and write our help centre, and we make an AI quiz generator for training courses, so weigh what I say about it with that in mind. Instructional designers like me tend to write the questions last, after the content is "done". I now write the check first, because the check is the part of the lesson that does the remembering.

What is a knowledge check?

A knowledge check is a short set of low-stakes questions placed inside or straight after a lesson, so learners recall what they just learned and get feedback before they move on. It is usually only a few questions, it is usually ungraded, and learners can usually try again.

You will also see it called a checkpoint, a quick check, a practice quiz, a review question or a check for understanding.

A knowledge check is a form of formative assessment: it informs the learning while it is still happening, instead of judging it at the end. Our guide to formative vs summative assessment covers that wider split. It also differs from a pre-training assessment, which finds out what people know before the training starts.

Knowledge check vs quiz

The two words get used interchangeably, but a knowledge check and a graded quiz do different jobs.

Because nothing rides on a knowledge check, you can show the right answer and the reason straight away without spoiling a test.

When the result contributes to certification, access or a compliance decision, define the required assessment and records. A quiz can check relevant knowledge; practical or regulated capabilities may also need observed evidence and an authorised assessor. Keep low-stakes practice as well as the required assessment.

Why knowledge checks work: the testing effect

Retrieving an answer from memory strengthens the memory more than reading the answer again. Psychologists call this the testing effect, or retrieval practice.

In 2006, Henry Roediger and Jeffrey Karpicke had students read short prose passages and then either reread them or take recall tests on them, with no feedback. When the final test came five minutes later, rereading won. On the delayed tests it reversed, and their paper in Psychological Science reports that prior testing produced substantially greater retention than studying, even though the students who reread felt more confident they would remember.

The confidence finding is the useful warning for anyone designing training: the version that feels smooth at the time is not the one people remember the next week. A 2017 meta-analysis by Adesope, Trevisan and Sundararajan pooled the research and found practice tests more beneficial for learning than restudying and every other comparison condition they looked at. For the wider set of retention methods, see our guide to knowledge retention.

A wrong answer also shows the learner a gap while the lesson is still in front of them, and shows you which part of the lesson is not landing.

Case study: Repeated testing for medical residents in St Louis

Laboratory studies of the testing effect usually give the final test a short time after learning. Douglas Larsen, Andrew Butler and Henry Roediger at Washington University in St Louis tested it in a real teaching setting and measured recall six months later. Six months is a more useful horizon for workplace training than a few days.

• What they did: Paediatric and emergency medicine residents attended an interactive teaching session on two topics, status epilepticus and myasthenia gravis. Each resident was randomised to take short-answer tests on one topic and study a review sheet on the other. The review sheet held the same information as the test answers. Testing or studying happened straight after the session and twice more at about two-week intervals, and residents got feedback after each test.
• Evidence and result: Forty residents completed the study (19 and 21 in the two groups). On a final test about six months later, Larsen and colleagues report in Medical Education that the tested topic scored 39% on average against 26% for the studied topic, a gap of 13 percentage points.
• Limitation: The sample was small and came from one teaching conference, and the final test measured recall of facts, not how the residents treated patients. Both scores were low after six months, so testing slowed forgetting rather than preventing it.
• Lesson: The same information, given as questions with feedback instead of as a review sheet, was remembered better half a year later. If your lesson ends with a summary slide, turn it into a few questions with feedback, then bring the same questions back twice over the following month.

Types of knowledge checks

Two useful ways to group checks are recall and application. Recall asks for a fact, term or step; application asks the learner to use knowledge in a situation. Match the balance to the job: some facts need reliable recall, and decisions need relevant cases. These are purposes, not an exhaustive taxonomy.

Either type can use any of these formats:

• Single answer: one correct option out of three. The easiest format to give feedback on.
• Multiple answer: tick every option that applies. Good for lists the learner must know in full.
• Matching pairs related items, such as terms and definitions. Ordering asks for a sequence. Drag and drop is an interaction either can use; it does not by itself show the learner can perform the task.
• Short answer: the learner types the answer. Harder to fake, and the closest to free recall.

A scenario is not a separate format: it is a short situation in the question stem, followed by one of the formats above, and it is how most application checks are written. True or false questions are quick to write, but a learner guessing at random gets about half of them right, so they tell you little on their own.

How to create a knowledge check (worked example)

Every step below is applied to one hypothetical lesson, built to show the method. It is part of a contact-centre induction, and it teaches new agents to verify a caller before discussing their account. The agents are in their first week and have not taken live calls yet. The risk is that an agent, trying to be helpful, gives account details to someone who is not the account holder.

The policy in the lesson is short. The caller gives their full name and date of birth, plus one more detail from the account: the postcode on file or the amount of their last payment. The agent never reads account details out for the caller to confirm. If the caller cannot pass, the agent does not discuss the account and offers a call back on the number already on file.

This is an invented policy for demonstrating question design, not a recommended identity-verification standard. Use the organisation's current approved process and security review. Calling a stored number alone does not establish identity, and account information stays protected until the required verification is complete.

Start from one objective

Write the objective as something the learner will do, then sample the decisions and likely errors it includes. One simple objective may need several questions.

• The objective: Before discussing an account, the agent verifies the caller with name, date of birth and one more account detail, and never prompts the answer.

If the objective is hard to state clearly, review its scope and the decisions it includes before choosing the checks. A long sentence alone does not establish that two lessons are needed. Our guide to SMART learning objectives with examples shows how to tighten one.

Ask for a decision, not a definition

A recall stem such as "Which details must the caller give?" tests whether the agent can recite the policy. A real call will put a person on the line with a reason why the policy does not quite fit. So write the stem as that moment, the same way scenario-based learning works at course scale.

• The question: A caller gives the right name and date of birth, then says, "I moved last month, so the postcode will be wrong." What do you do next?

Write three plausible options

Research by Andrew Butler at Washington University, summarised in the university's own report on how to write better multiple-choice tests, recommends three plausible response options, simple question formats, and no "all of the above" or "none of the above". Each wrong option should be a mistake a real new agent might make.

• Options: (a) Ask for the amount of their last payment. (b) Go ahead: name and date of birth are enough for a quick question. (c) Read out the postcode on file so they can confirm it.

Option (a) is correct. Option (b) is the shortcut a busy agent takes. Option (c) is the helpful move that breaks the policy. Compare a weaker version of the same item:

• Weak options: (a) Ask for the amount of their last payment, or another detail listed in the policy. (b) Hang up. (c) Ask them their favourite colour.

The joke options test nothing about the job, and the correct answer is the longest one, which a learner can spot without knowing the policy.

Write feedback for every option

Feedback is what makes a multiple-choice check safe. The wrong options, which researchers call lures, are written to be believable, and a learner who picks one can walk away remembering it as true. In Butler and Roediger's 2008 study in Memory and Cognition, picking lures led students to acquire false knowledge, and both immediate and delayed feedback reduced it. The authors conclude that educators should provide feedback when using multiple-choice tests.

So each option gets its own line, and each wrong line names the mistake behind it:

• Feedback (a): "Right. The postcode is one option for the third check, not the only one. The last payment amount is the other."
• Feedback (b): "Not yet. Name and date of birth are often known to family members or easy to find online. You need a third detail from the account."
• Feedback (c): "Never read details out for the caller to confirm. That hands the answer to anyone on the line."

Writing these lines also tests the item. If you cannot say in one sentence why a wrong option is wrong, the option is either correct or too vague, and it needs replacing.

Feedback is also the part that course briefs leave out. We classified a sample of the requests people typed into the Coursebox course builder over 90 days to see how often they ask for it.

Most briefs in this sampled set that asked for questions did not say what should happen when a learner gets one wrong. When you brief a course builder or a colleague, ask for the feedback lines in the same sentence as the questions.

Keep it to a few questions

Keep the check focused, but use enough questions to cover the important decisions and likely errors. Extra items can sample different cases; remove items that add neither evidence nor useful practice. Here three questions rehearse the example policy, without claiming to establish secure call handling.

• Item 2, the question: Which of these can be the third detail? Tick all that apply.
• Options: Postcode on file. Amount of last payment. Caller's email address.
• Feedback: Postcode on file: "Yes, if it matches the account." Last payment: "Yes, the exact amount of their most recent payment." Email address: "Not on our list. Email addresses are easy to find, so they do not prove who is calling."
• Item 3, the question: The caller cannot give the postcode or the last payment amount and is getting frustrated. What do you do?
• Options: (a) Offer to call them back on the number on file. (b) Ask their date of birth again. (c) Tell them which detail they got wrong so they can try again.
• Feedback: (a) "Right under this example policy: offer the stored-number callback route without disclosing account details. Complete the approved verification before discussing the account; the callback alone is not proof of identity." (b) "Repeating an already supplied detail does not complete the check." (c) "Do not supply the answer or identify which detail failed. Follow the approved next step politely."

If you draft items with an AI assistant, our ChatGPT prompts for quiz writing ask for plausible wrong options and feedback, which you then check against the policy.

Where to place knowledge checks in an e-learning course

Put a check after each chunk of the lesson, not only at the end. In an online lesson nobody notices when the learner drifts, and a check every few minutes gives them a reason to pay attention to the next chunk.

Karl Szpunar, Novall Khan and Daniel Schacter tested this with recorded online lectures at Harvard. Their 2013 study in PNAS found that breaking the lecture up with short memory tests reduced mind wandering, increased note-taking and improved learning, compared with the same lecture without the tests. The tested group also reported less anxiety about the final cumulative test.

For the verification lesson, that means a check after the part on the three required details and another after the part on what to do when a caller fails, rather than all three items at the end. Then bring the same decisions back a few days or weeks later. Our guide to spaced repetition covers how to schedule that.

Reading the results

An ungraded check still produces data: who picked which option. Read it item by item, and look at the pattern rather than the average.

• Most learners miss one item: first inspect wording, accessibility and whether more than one answer is defensible. Then check the explanation, prior knowledge and opportunity to practise. The pattern is a reason to investigate, not proof of one cause.
• Most wrong answers land on one option: it may represent a common misconception or a misleading item. Ask a few learners to explain their choice before changing the lesson. In this example, choosing the postcode read-out option would be a reason to revisit the no-prompting rule.
• Almost everyone gets it right first time: The item may be too easy, or answerable from the stem alone. Check whether the correct option is the longest or the most detailed.

None of these needs a cut-off number. Look at which option people chose, and ask whether the item or the lesson caused it.

Common knowledge check mistakes

• Trivia questions: asking for a date, a policy number or a definition the job never uses.
• All of the above: a learner who spots two right options gets the item without knowing the third.
• The longest answer is correct: careful writers make the correct option the most qualified one, and learners learn to pick it.
• No feedback: a wrong answer with only a red cross can stay in memory as the right one.
• Every check at the end: one long quiz after the content leaves nothing to break up the lesson itself.

Building a knowledge check in Coursebox

I wanted to see whether an AI course builder would produce a knowledge check that does this job, so I ran the caller verification lesson through the Coursebox course agent in a test workspace.

What I asked for. I pasted the one-page verification policy and asked: "Build a 10-minute induction lesson on verifying a caller before discussing their account, from this one-page policy. After the lesson add a knowledge check: 3 questions, scenario-based, 3 options each, feedback on every option, not graded."

What it produced. The agent did not ask any clarifying questions. It planned one section with two pages, at 15 minutes rather than the 10 I asked for: a knowledge page called "Verify the Caller First" and a practice page called "Knowledge Check: Make the Safe Call". On content, the check did what I asked: three decision points, three options each, a feedback line on every option and no "all of the above". The stems were decisions, for example: "The caller has given their full name and date of birth but cannot provide the postcode. They ask you to read it out. What should you do?" The read-out mistake was there as a serious option, with the feedback "Reading out a postcode discloses an account detail. The caller must supply the postcode themselves."

What was wrong. The check was built as a custom HTML widget, a block of code with its own buttons, not as a Coursebox quiz block. The page used a custom widget rather than the native quiz settings. I did not verify that it saved answers or sent them to course reporting, so I could not rely on it for a manager's item-level record. The correct answer was also the first option in all three questions, with no shuffle, so a learner who noticed could pass without reading.

The fix, and how I checked it. I sent the AI editor one message asking it to replace the widget with a native quiz block, keep the three scenarios and three options, vary the position of the correct answer, put the reason each wrong option is wrong into each question's hint, and turn grading off. Its reply said it had done all of that, with the correct answers first, second and third. I reloaded the page and checked the page itself. The widget was gone, and the page held a native quiz block with three single-answer questions of three options each, with "Graded quiz" off. The correct answers sat first, third and third, not as the reply said. Two of the three hints pointed at the wrong options: in question 2 the hint called option 3 wrong when option 3 was the correct answer. The hints also refer to options by number, which breaks if "Shuffle Answer Order" is turned on. A multiple-choice question in a Coursebox quiz block takes one hint, not a line per option, so the per-option feedback the widget had became one hint per question, and those hints need rewriting by hand before any learner sees them. I have not yet run the check with test learners. After any AI edit, read the page itself instead of trusting the confirmation message.

Putting knowledge checks to work

Grade the final quiz if you need a record, and leave the checks inside the lesson ungraded so people answer honestly and read the feedback.

Pick one lesson you already run. Write its objective in one sentence, then write one decision question with three plausible options and a feedback line for each. Put it straight after the part of the lesson it tests, and run it with two or three learners before launch. Ask why they chose each answer; their mistakes can reveal unclear wording, missing explanation or a need for more practice. The planner above has the caller verification item filled in as a model.