TL;DR
The researchers tested 3 prompting strategies for generating quiz questions: a detailed prompt that describes the cognitive level and lists action verbs, a minimal prompt with no specifics, and a persona prompt ("seasoned professor"). The goal: produce questions that land exactly on the levels of Bloom's taxonomy (Knowledge, Application, Analysis).
The main finding: without explicit instructions, an LLM systematically misses the cognitive level. The minimal prompt "make a Knowledge-level question" produces a Comprehension-level question 60% of the time. The persona prompt is even worse: 60% misses. Meanwhile every question is clear and relevant (5.0 out of 5.0), just at the wrong cognitive level. The problem is that the model doesn't grasp the difference between "recall a fact" and "explain a concept" unless that difference is spelled out.
The detailed prompt works: add the definition of the Bloom level plus concrete verbs ("define", "identify" for Knowledge; "apply", "demonstrate" for Application) and accuracy jumps to 96%. The model gets clear criteria instead of abstract level names and generates questions at exactly the requested cognitive depth.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
A SINGLE-STEP PROMPT with explicit structure:
PROMPT:
Create a question about {topic} at the {Bloom level} level.
This level means:
{level definition}
Use these action verbs:
{list of verbs for this level}
The question must be a complete sentence.Result: One question that matches the requested cognitive level exactly.
Example in Practice
⚠️ Important: The method works for structured tasks where hitting the right cognitive level matters (training, testing, course materials). Not for creative or open-ended questions.
Task: You're a marketer putting together an internal quiz for interns on the basics of search advertising. You need Application-level questions, so they actually apply what they know instead of reciting definitions. Your pain point: when you type "create a question that tests applying knowledge", ChatGPT gives you either a trivial "what is CTR?" or a three-paragraph analytical case study.
Prompt:
Create a question for an intern quiz about setting up targeting in Google Ads.
Level: Application (applying knowledge).
This level means: using information in a new situation, solving
practical problems, applying rules and concepts.
Use these action verbs: apply, use, demonstrate,
carry out, solve, create, choose.
The question must be a complete sentence, and the scenario must be concrete.Result:
The model will produce a question like: "You're launching ads for a sports nutrition brand in Chicago. The target audience is men aged 25-35 who work out at the gym. Apply location and interest targeting settings in Google Ads for this campaign. Describe which parameters you'd choose and why."
Instead of: "What is geotargeting?" (Knowledge) or "Analyze the effectiveness of 5 targeting strategies across different audience segments" (Analysis).
Why It Works
LLMs are bad at abstract cognitive levels. When you write "Knowledge level", the model doesn't know where exactly "recall" ends and "explain" begins. It just sees the word "Knowledge" with no context and may generate either one. Result: a miss 40-60% of the time.
LLMs are great with concrete instructions. Give it a definition ("use information in a new situation") plus marker verbs ("apply", "demonstrate", "solve"), and the model gets a clear pattern to generate from. It builds the question around those verbs and stays focused on the right cognitive depth.
A persona without criteria doesn't work. "You are a seasoned professor" is context, not a specification. The model reads the persona as "make it harder" or "sound more expert", but it doesn't pick up the exact cognitive level. Effect: questions come out harder than needed (60% misses).
Levers for tuning the prompt:
- Level definition: remove it and accuracy drops from 96% to 60%. This is the model's main anchor.
- Verb list: swap in your own if you want a specific format (e.g., for medicine: "diagnose", "prescribe"; for law: "classify under the statute", "apply the rule").
- Task context: add "the scenario must be concrete" if you want case studies; drop it if you want abstract questions.
Prompt Template
Create a question about {topic} at the {level} level.
This level means: {level definition from the table below}.
Use these action verbs: {verbs from the table below}.
The question must be a complete sentence, {additional requirements}.Bloom's levels table:
| Level | Definition | Action verbs |
|---|---|---|
| Knowledge | Recall facts, terms, basic concepts | define, name, list, memorize, recognize, find |
| Comprehension | Explain ideas, interpret information | explain, describe, interpret, summarize, compare |
| Application | Use information in a new situation, solve problems | apply, use, demonstrate, carry out, solve, create |
| Analysis | Break information into parts, find connections and causes | analyze, break down, compare, examine, identify, determine motives |
| Synthesis | Combine parts into a new whole, generate new ideas | develop, create, design, build, propose, formulate |
| Evaluation | Judge the value of information, justify a decision | evaluate, justify, argue, critique, defend, judge |
What to plug in: {topic} is your subject area, {level} comes from the table, {additional requirements} is something like "the scenario must be concrete" or "the question should be 2-3 sentences long".
🚀 Quick start: paste this into the chat:
Here's a template for generating questions at a precise cognitive level.
Adapt it to my task: [describe what you need].
Ask me questions to fill in the fields.
[paste the template above]The LLM will ask: what topic, what level, any format requirements, because it needs those parameters to fill in the structure. It will take the pattern from the template and build a working prompt for your task.
Limitations
⚠️ Narrow domain: Only Computer Science was tested, and only 3 Bloom levels (Knowledge, Application, Analysis). For the humanities or other levels you may need to adapt the verbs and definitions.
⚠️ One model: The experiment ran on Gemini 2.0 Flash Lite. Other models (GPT-4, Claude) may react differently to persona or minimal prompts, but the "detailed instructions win" principle is universal.
⚠️ Not for creative tasks: The method is built for structured content with clear cognitive goals. If you need creativity or open-ended questions, detailed criteria can limit variety.
How They Tested It
The team took 5 concepts from Computer Science (AI, Data Structures, Cybersecurity, Web Development, Cloud Computing) and, for each one, generated questions at 3 Bloom levels (Knowledge, Application, Analysis) with three different prompts. That's 135 questions in total (5 concepts × 3 levels × 3 questions × 3 prompts).
Every question went through a DistilBERT classifier (a model trained to detect the Bloom level), and they computed the Match Rate: the share of questions where the automatically detected level matched the one requested in the prompt. On top of that, 45 questions (a third of the sample) were rated by hand on a 1-5 scale for clarity, topic relevance, and subjective level alignment.
Why did the results come out this way? The detailed prompt (A) hit 96% accuracy: the model got explicit criteria and didn't have to guess. The minimal prompt (B) scored 60%: the model filled in the blanks itself and often overshot the difficulty ("Knowledge" → "Comprehension"). The persona prompt (C) scored 40% (the worst!): the "seasoned professor" nudged the model toward harder questions instead of hitting the level precisely.
The surprising part: All three variants scored 5.0 out of 5.0 on clarity and relevance. So the questions were understandable and on topic, just at the wrong cognitive level. It shows that text quality ≠ pedagogical precision. A model can write something polished and correct and still miss the point of the assignment.
The stats back it up: ANOVA showed significant differences between the variants (F(2,132)=25.00, p<0.001). Post-hoc tests: variant A is statistically better than B and C (p < 0.001), and variant B beats C (p < 0.05). This isn't noise; the gap is real and stable.
Original Prompts from the Paper
Context: The researchers used the generation_model.predict function with temperature=0.7 for all generations. Here are the three prompts verbatim:
Variant A: Baseline/Explicit Bloom Prompt
Generate a question about '{concept}' for the '{level}' level of Bloom's Taxonomy.
The question should specifically align with the '{level}' level, which involves the
following characteristics: {level_descriptions[level]}. Use action verbs like
{', '.join(bloom_verbs[level])}. Ensure the question is a complete, clear sentence
and relevant to the concept. Avoid titles or headings.Variant B: Simpler/Implicit Prompt
Generate a question about '{concept}' at the '{level}' level of Bloom's Taxonomy.
The question should be a complete, clear sentence.Variant C: Persona-Based Prompt
As a seasoned computer science professor designing an exam, generate a question
about '{concept}' for a university student. The question should be at the '{level}'
level of Bloom's Taxonomy and effectively assess their understanding. Ensure the
question is a complete, clear sentence.Where {concept} is the topic (e.g., "Artificial Intelligence"), {level} is the Bloom level (e.g., "Application"), {level_descriptions[level]} is a text description of the level, and {bloom_verbs[level]} is the list of verbs for that level.
Resources
Yaacoub, A., Da-Rugna, J., & Assaghir, Z. (2025). Lightweight Prompt Engineering for Cognitive Alignment in Educational AI: A OneClickQuiz Case Study. Proceedings of the 36th Central European Conference on Information and Intelligent Systems, 489-496.
Related work by the authors:
- Assessing AI-Generated Questions' Alignment with Cognitive Frameworks (arXiv:2504.14232)
- Cognitive Depth Enhancement via SOLO Taxonomy (Springer, 2025)
- OneClickQuiz: Instant AI-Driven Quiz Generation in Moodle (SEED 2024)
Antoun Yaacoub, Jérôme Da-Rugna: ESIEA, Paris Zainab Assaghir: Lebanese University, Beirut