Assessment content is the part of e-learning production nobody wants to talk about. The videos get the attention, the graphics get the budget, and then someone realises there are 400 lessons that each need three quiz questions — stem, four options, correct answer, brief explanation — and the launch is in six weeks.
For Dan, head of content at an online professional learning platform, this was a familiar crunch. The platform had grown to 400 lessons across eight subject areas, and the quiz coverage was patchy: some lessons had good questions, some had thin ones written under deadline pressure, and about a third had none at all. A proper rebuild of the question bank was overdue.
Manual question writing runs at roughly 15–20 minutes per question when done carefully — accounting for writing the stem, drafting plausible distractors, checking the correct answer is unambiguous, and writing an explanation that actually teaches something. At that rate, 1,200 questions was a 300–400 hour project. Six weeks, one writer, no slack.
The batch ran overnight. By morning he had a draft question bank to review.
Building the spreadsheet
Dan's lesson content was in a CMS. He exported the relevant fields to a spreadsheet — one row per lesson.
| lesson_id | lesson_title | subject_area | difficulty | lesson_summary |
|---|---|---|---|---|
| L0042 | Understanding Cash Flow Statements | Finance | Intermediate | Covers the three sections of a cash flow statement (operating, investing, financing), how to read net cash position, and the difference between profit and cash flow. Key concept: a profitable business can still run out of cash. |
| L0117 | Negotiation: BATNA and Reservation Price | Business Skills | Advanced | Defines Best Alternative to a Negotiated Agreement (BATNA) and reservation price. Covers how to calculate your BATNA, why revealing it weakens your position, and how to anchor effectively in opening offers. |
The lesson_summary column did the most work. A good summary — even two or three sentences — gave the model enough to write questions that were genuinely grounded in the lesson content rather than generic questions about the topic. The difficulty column controlled the cognitive level of the questions: an intermediate lesson should test application, an advanced lesson should test analysis.
For lessons where the CMS summary was thin, Dan spent ten minutes writing a better one before running the batch. Time well spent — the output quality correlated directly with the quality of the input summary.
Writing the prompt
The most important design decision was the output format. Quiz questions need to import cleanly into an LMS — each field (stem, options, answer, explanation) has to land in its own column or be easily split. The prompt locked this down.
Prompt used:
You are an instructional designer writing multiple-choice assessment questions for a professional e-learning platform. Generate exactly 3 quiz questions based on the lesson described below.
Each question must follow this exact format — repeat it three times, labelled Q1, Q2, Q3:
Q[N]_STEM: [the question — one clear, unambiguous sentence]
Q[N]_A: [option A]
Q[N]_B: [option B]
Q[N]_C: [option C]
Q[N]_D: [option D]
Q[N]_CORRECT: [A / B / C / D]
Q[N]_EXPLANATION: [one sentence explaining why the correct answer is right and the key misconception the wrong answers represent]
Question quality rules:
— Base all questions on the lesson_summary content only — do not introduce concepts not covered in the lesson
— Vary cognitive level across the three questions: one recall, one application, one analysis
— Distractors must be plausible — wrong answers that a learner who half-understood the lesson might genuinely choose
— Avoid "all of the above" and "none of the above" options
— Do not use the word "not" in the stem — reframe as a positive question
— Difficulty level for this lesson: {difficulty} — calibrate accordingly
— Each option should be roughly the same length — avoid giving away the answer through option length
The rule about option length is one those things that sounds minor but makes a real difference. If the correct answer is always the longest option, learners learn to game the test rather than the content.
Running the batch
400 rows on Gemini 2.5 Flash. The inputs are structured and the output format is tightly constrained — Flash handles this class of task reliably. At 400 lessons the cost difference between Flash and Pro is significant; at 1,200 questions the output quality from Flash was good enough that upgrading wasn't necessary.
The batch completed in about an hour. Total cost: £1.40.
What the output looked like
For the cash flow lesson:
Q1_STEM: Which section of a cash flow statement records money received from customers and paid to suppliers?
Q1_A: Investing activities
Q1_B: Financing activities
Q1_C: Operating activities
Q1_D: Retained earnings
Q1_CORRECT: C
Q1_EXPLANATION: Operating activities cover day-to-day cash flows from the core business; investing and financing activities cover assets and capital respectively, while retained earnings appears on the balance sheet, not the cash flow statement.
Q3_STEM: A company reports a net profit of £200,000 for the year but ends the period with a negative cash balance. Which of the following best explains how this is possible?
Q3_A: The company's accounting records contain errors
Q3_B: Revenue was recognised before cash was received from customers
Q3_C: The company paid its taxes twice in the same period
Q3_D: Retained earnings were incorrectly calculated
Q3_CORRECT: B
Q3_EXPLANATION: Accrual accounting recognises revenue when earned, not when cash is received — so a business can be profitable on paper while cash is tied up in unpaid receivables, making cash flow negative despite reported profit.
Q1 is recall. Q3 is analysis — exactly the cognitive spread the prompt asked for.
The review process
Dan reviewed all 1,200 questions over two days, working lesson by lesson. His pass rate — questions accepted without edit — was about 78%. The remaining 22% needed one of three types of fix:
- Distractor quality (most common): One of the wrong options was obviously wrong rather than plausibly wrong. A quick edit to make it more tempting. The prompt addressed this but the model occasionally slipped, especially on lessons where the subject matter had fewer natural misconceptions to draw on.
- Stem ambiguity (occasional): A question that could technically be answered two ways depending on interpretation. Rewritten to be more precise.
- Content drift (rare): A question that was technically correct for the topic but went beyond what the lesson summary covered. Replaced with something grounded in the actual lesson content.
At 78% acceptance, the batch saved roughly 235 hours of writing time — everything beyond the two-day review pass. The questions it produced weren't perfect, but they were a solid first draft: correctly formatted, well-structured, and grounded in the lesson content. That's exactly what makes a good draft valuable.
Importing into the LMS
The structured output format — one field per line, clearly labelled — made the import straightforward. Dan used a simple formula to split each row's AI responses into 14 separate columns (stem, A, B, C, D, correct answer, explanation × 3 questions), then mapped those columns to his LMS's bulk import template. The full import took about an hour.