After this lesson you can ask for an answer in the shape and reading level you need.
About 7 minutes, or 13 with the Deep layers
Reading depth for every section
In short
A request can name the shape you want: a table, its columns, and how long each sentence may be.
In this lesson's recordings, asking for a shape got that shape. Count and read the answer to check it.
Saying a rule again at the end of a long message changed nothing we could count in these runs.
Reading depth forListen to
One task, two requests
Here is one task: compare three ways to revise for an exam. We sent it to two models in two ways.
The first request was only that sentence. The second also named a shape: a table, four columns, plain words, and sentences under 15 words. The columns were method, what it helps with, time needed, and one tip.
The models are Claude Sonnet 5.5, made by Anthropic, and Ministral 3 14B, made by Mistral AI. Each request went to each model three times. Every answer was recorded on 3 October 2026.
The request said what to compare and nothing about the answer’s shape. So the shape of each reply was the model’s own choice. That is what the first request tests.
Both models were called through each maker’s API (a way for one program to send a request to another), with no tools and no system prompt, and the temperature was not set for either. Ministral 3 14B’s model card in the recording lists a default temperature of 0.3. The reply limit was set to 4,000 tokens for both.
What a model does with a request is a fact about that model on that day. Another model, or another day, can answer in another shape.
This lesson looks at shape. It does not check whether the revision advice in the answers is right.
Ask for the shape
You can ask for the content only. But you can also name the shape. You can name the layout, such as a table or a list. You can name the columns. You can name how long a sentence may be.
A shape you can use as it comes saves you from reshaping the answer by hand.
Below, the activity asks you to predict the shape of the answers before you see them. It shows every run, with its model and date. The answers are there.
PromptSage is a prompt framework by Emanuel Covasa, written up as a paper that is not yet published. This lesson names the shape of an answer; PromptSage also names the parts of the prompt that asks for it. It splits a prompt into seven labelled layers, each written between XML tags, such as <identity> and </identity>. The opening and closing tags mark where each part starts and ends.
Disclosure: PromptSage was created by Emanuel Covasa, who also created this site. It is one worked framework, not the standard.
Anthropic says XML tags help Claude read a prompt that mixes instructions with other content, such as context, examples and inputs. It says wrapping each kind of content in its own tag “reduces misinterpretation” (read on 4 October 2026). This is its advice for Claude, not a guarantee. You can learn more about XML tags in Anthropic’s guide (opens in a new tab), a free page on a different website, written for developers.
The activity has one more part, for the Deep layer: a layer explorer. It shows six to eight rules from an invented study-helper prompt. For each rule, you predict which of PromptSage’s seven layers it belongs in. Then you see the same prompt built layer by layer. That prompt is not a recording. We wrote it for this lesson, and no model answered it.
A request can also fix a shape by example. Lesson 2.4 shows that. This lesson uses a description of the shape instead.
Lesson 2.1 shows the same idea in one line: say what shape the answer should take.
Try it: Shape the answer
We sent two requests to two AI models, three times each. The requests are invented.
Request A
Compare three ways to revise for an exam.
Request B
Compare three ways to revise for an exam in a table with the columns: method, what it helps with, time needed, one tip. Use plain words and sentences under 15 words.
What all the runs showed
With a shape asked for, all six runs gave a table with the four named columns. With no shape asked for, all six gave a long answer with headings, lists and a table.
Every sentence in the six tables was under 15 words, by our count. Ministral 3 14B added one line before its table in all three runs.
In the long message, all twelve runs gave three top-level bullets. Saying the rule twice changed nothing we could count.
Claude Sonnet 5.5 sometimes added smaller bullets inside them. That happened in one run of the first message, and in three of the second.
The two AIs and their settings
Claude Sonnet 5.5
claude-sonnet-5-5, on Anthropic Messages API. Recorded on 3 October 2026.
Longest answer, in tokens: 4000
Temperature: not sent, and the value the service used was not recorded
Tools: no tools
Instructions from us behind the message: none
Ministral 3 14B
ministral-14b-2512, on Mistral AI API. Recorded on 3 October 2026.
Model name sent: ministral-14b-2512
Longest answer, in tokens: 4000
The model's default temperature, read from its model list: 0.3
Temperature: not sent, and the value the service used was not recorded
Tools: no tools
Instructions from us behind the message: none
Request A: no shape asked for
These runs are written out in full at the end of this activity.
Request B: the shape asked for
These runs are written out in full at the end of this activity.
The rule once, and the rule twice
These are Claude Sonnet 5.5's answers to two long messages. One has the rule once. The other also has it as the last line. They start with two answers that have smaller bullets inside the three.
Hello. I need some advice on running a study group. I am a second-year student and I have agreed to organise one for my statistics module. There are eight of us in the group, but only five usually come. We meet in the library on Tuesday evenings from six until eight, because that is the only time the room is free. The module has a weekly problem sheet, and most of us get stuck on the same questions. Two members are very quiet and rarely speak unless they are asked. One member talks for a long time and often moves the conversation away from the problem sheet. Last week we spent forty minutes talking about the lecturer's slides and finished only two questions.
Answer in exactly three bullet points.
I would like the group to feel friendly, because several of us are nervous about the exam in January. I do not want it to feel like a second lecture, and I do not want to be bossy, because the others are my friends. Some members prefer to work alone first and then compare answers, while others like to start together at the whiteboard. We have no tutor with us, but our lecturer has said she will answer a short email each week. We also have a shared online folder where we keep notes. Our group chat is busy, but most messages are about when to meet rather than about the work. A few people have said they feel they are not learning much, and one person has already said they may stop coming. Please give me advice on how to run the sessions so that everyone takes part and we finish more of the problem sheet. Thank you, and sorry that this is a long message.
These runs are written out in full at the end of this activity.
Show all runs
Request A: no shape asked for
Claude Sonnet 5.5 Run 1 Recorded 3 October 2026
No shape was asked for. A long answer with headings, lists and a table, 344 words.
Three ways to revise for an exam
1. Re-reading and highlighting notes
How it works: You go over your notes or textbook and mark key points.
Pros: Low effort, easy to start, and good for a first pass to get familiar with the material.
Cons: It's passive. It creates a feeling of familiarity that's easily mistaken for real understanding, and research consistently rates it among the least effective methods.
Best for: An initial overview, or a quick refresher the day before.
Read the full recorded answer, Claude Sonnet 5.5, run 1
2. Active recall (flashcards, self-quizzing, "blank page" recall)
How it works: You test yourself on the material without looking at it, then check your answers.
Pros: Strongly supported by research. It exposes gaps in your knowledge, strengthens memory, and works especially well when spaced over several days or weeks.
Cons: It feels harder and slower, and making good flashcards takes time. It can encourage isolated facts over deeper understanding if you're not careful.
Best for: Fact-heavy subjects like biology, history, languages, and definitions.
3. Practice papers and past questions
How it works: You answer real exam-style questions, ideally timed, then mark them against the mark scheme.
Pros: It trains the actual skill being assessed, including applying knowledge, managing time, and structuring answers. It also shows you the exam's patterns and your weak spots.
Cons: You need some knowledge first, or it can be demoralising. Past papers are limited in number, and marking your own work takes discipline.
Best for: Maths, sciences, essay-based subjects, and the final weeks before the exam.
Side-by-side
Re-reading
Active recall
Practice papers
Effort
Low
Medium–high
High
Effectiveness
Low
High
High
Builds exam technique
No
Slightly
Yes
Shows weak spots
Poorly
Well
Very well
Best stage
Early
Throughout
Late
Recommendation
These methods work best in sequence. Start with a quick read-through to get the material in your head. Then move to active recall, spaced across several days. Finish with timed practice papers, and use your mistakes to decide what to revisit.
If you tell me your subject and how much time you have, I can suggest a more specific plan.
Claude Sonnet 5.5 Run 2 Recorded 3 October 2026
No shape was asked for. A long answer with headings, lists and a table, 352 words.
Three Ways to Revise: A Comparison
I've compared rereading and highlighting (the common default) against two techniques with strong research support.
1. Rereading and Highlighting Notes
How it works: You read through your notes or textbook again, marking key points.
Pros: Easy and low effort, feels productive, and is a good first pass to get familiar with material.
Cons: It creates an "illusion of knowing," because material looks familiar without being retrievable from memory. It's slow and gives poor long-term retention.
Best for: An initial overview, or a final skim the night before.
Read the full recorded answer, Claude Sonnet 5.5, run 2
2. Active Recall (Practice Testing)
How it works: You test yourself without looking at the material, using past papers, flashcards, or blank-page "brain dumps," then check your answers.
Pros: It is among the most effective methods in the research. It shows you exactly what you don't know, and it mimics exam conditions.
Cons: It feels harder and more frustrating, and it requires preparing questions or finding past papers.
Best for: Fact-heavy subjects (sciences, history, languages) and problem-based ones (maths, physics).
3. Spaced Repetition
How it works: You review material at increasing intervals (e.g., after 1 day, 3 days, 1 week, 3 weeks) rather than cramming. Apps like Anki automate this.
Pros: It greatly improves long-term retention, reduces last-minute panic, and works well combined with active recall.
Cons: It requires planning and starting early, so it's no use if the exam is tomorrow. Building flashcard decks takes time.
Best for: Cumulative subjects and anything with a lot to memorize over weeks or months.
Side-by-Side
Rereading
Active Recall
Spaced Repetition
Effort
Low
High
Medium
Effectiveness
Low
High
High
Time needed
Moderate
Moderate
Spread over weeks
Feels like it works?
Yes (misleadingly)
Less so (but it does)
Moderately
Works for last-minute?
Somewhat
Yes
No
Recommendation
Combine the last two: use active recall as your main method, and space those sessions over days and weeks. Keep rereading to a minimum, mainly for first exposure to the material. If time is short, prioritize past papers and self-testing.
If you tell me the subject and how long you have, I can suggest a specific plan.
Claude Sonnet 5.5 Run 3 Recorded 3 October 2026
No shape was asked for. A long answer with headings, lists and a table, 323 words.
Three Ways to Revise for an Exam
1. Re-reading and highlighting notes (passive review)
How it works: You go back over your notes or textbook and mark key points.
Pros: Easy and low-effort, and good for a first overview of the material.
Cons: It creates a false sense of familiarity. Recognizing information is not the same as being able to recall it. Research consistently ranks it among the least effective methods.
Best for: The very start of revision, or a quick skim the day before.
Read the full recorded answer, Claude Sonnet 5.5, run 3
2. Practice testing / active recall (past papers, flashcards, self-quizzing)
How it works: You retrieve information from memory without looking, then check your answers.
Pros: It is among the most evidence-backed methods. It shows you exactly what you don't know, builds strong memory, and trains you for exam conditions and timing.
Cons: It feels harder and more frustrating, and it needs good materials such as past papers or well-made questions.
Best for: Nearly every subject, especially those with factual content, problem-solving, or essay questions.
3. Spaced repetition (spreading study over time)
How it works: You revisit topics at increasing intervals (for example after 1 day, 3 days, then a week) instead of cramming.
Pros: It greatly improves long-term retention and reduces last-minute stress. It combines well with flashcards or apps like Anki.
Cons: It requires planning and starting early. It's less useful if the exam is only a day or two away.
Best for: Large amounts of content, such as languages, sciences, and history.
Quick comparison
Method
Effort
Effectiveness
Time needed
Best stage
Re-reading/highlighting
Low
Low
Moderate
Early overview
Practice testing
High
High
Moderate
Throughout, especially later
Spaced repetition
Medium
High
Spread over weeks
Start early, continue until the exam
Recommendation
The strongest approach combines the second and third methods. Use spaced repetition to schedule your revision, and use active recall during each session. Keep re-reading to a minimum, mainly for getting familiar with the material at the start.
Ministral 3 14B Run 1 Recorded 3 October 2026
No shape was asked for. A long answer with headings, lists and a table, 720 words.
Same run, a second claim marked.
Revising for an exam can be approached in many ways, each with its own strengths and weaknesses. Here are three effective revision methods, compared in terms of efficiency, retention, engagement, and suitability for different learners:
1. Active Recall (Spaced Repetition + Self-Testing)
How it works:
Instead of passively rereading notes, you actively retrieve information from memory (e.g., flashcards, practice questions, explaining concepts aloud).
Uses spaced repetition (reviewing material at increasing intervals) to strengthen memory retention.
Tools: Anki, Quizlet, self-made questions, past exam papers.
Pros: ✅ Best for long-term retention – Forces your brain to work harder, improving memory strength. ✅ Identifies weak spots – You quickly realize what you don’t know and can focus on those areas. ✅ Flexible & portable – Can be done anywhere (e.g., flashcards on the bus). ✅ Reduces illusion of knowing – Rereading makes you feel like you know the material, but active recall reveals true understanding.
Cons: ❌ Time-consuming upfront – Requires effort to create high-quality questions/flashcards. ❌ Can feel tedious – Less engaging than some other methods. ❌ Not ideal for big-picture subjects (e.g., history timelines) unless structured well.
Best for:
Fact-based subjects (e.g., biology definitions, math formulas, vocabulary).
Students who learn by doing (kinesthetic/active learners).
Cramming last-minute (if done with spaced repetition over time).
Example: Instead of reading your psychology notes, write down all the key theories and then test yourself without looking. Use Anki to schedule reviews.
2. Feynman Technique (Explain Like You’re Teaching)
How it works:
Choose a concept and explain it in simple terms as if teaching a 5-year-old.
Identify gaps – If you struggle to explain something clearly, go back to the source material.
Simplify & organize – Use analogies, diagrams, or examples to make it easier.
Review & refine – Repeat the process until you can explain it flawlessly.
Pros: ✅ Deepens understanding – Forces you to break down complex ideas into fundamental principles. ✅ Improves communication skills – Helps you articulate knowledge clearly (useful for exams with explanations). ✅ Great for conceptual subjects (e.g., physics, philosophy, economics). ✅ Reduces memorization – Encourages true comprehension over rote learning.
Cons: ❌ Time-intensive – Not efficient for memorizing large amounts of facts quickly. ❌ Requires strong foundational knowledge – Hard to apply if you don’t understand the topic well. ❌ Less structured – Doesn’t provide a clear revision schedule.
Best for:
Conceptual or theoretical subjects (e.g., calculus, ethics, political science).
Students who learn by teaching (visual/auditory learners).
Preparing for essay-based exams where understanding is key.
Read the full recorded answer, Ministral 3 14B, run 1
Example: If revising for a chemistry exam, try explaining how photosynthesis works without looking at your notes, then refine your explanation.
3. Interleaving (Mixing Topics & Problem Types)
How it works: Instead of blocking (studying one topic for hours at a time), you mix different subjects or problem types in a single session.
Example: Alternate between history dates, math problems, and language grammar in one study session.
Pros: ✅ Improves problem-solving skills – Helps the brain recognize patterns and apply knowledge flexibly. ✅ Reduces the "curse of knowledge" – Prevents overconfidence in one area while neglecting others. ✅ Mimics exam conditions – Many exams test a mix of topics, so interleaving prepares you better. ✅ Keeps revision engaging – Prevents boredom from repetitive studying.
Cons: ❌ Harder to master individual topics deeply – May feel less efficient for pure memorization. ❌ Requires good organization – Need a structured plan to avoid chaos. ❌ Not ideal for last-minute cramming – Works best when spread over weeks.
Best for:
Subjects with multiple sub-topics (e.g., mixed math problems, literature with different authors).
Students who struggle with applying knowledge (e.g., switching between algebra and geometry).
Long-term revision schedules (not just a few days before the exam).
Example: Instead of studying only Shakespeare for 2 hours, spend 30 mins on Shakespeare, 30 mins on Renaissance art, and 30 mins on medieval history in one session.
Comparison Table
Method
Best For
Retention
Engagement
Time Efficiency
Exam Preparation
Active Recall
Facts, definitions, memorization
⭐⭐⭐⭐⭐
⭐⭐⭐
⭐⭐⭐⭐ (long-term)
⭐⭐⭐⭐⭐ (practice questions)
Feynman Technique
Conceptual understanding, essays
⭐⭐⭐⭐
⭐⭐⭐⭐
⭐⭐ (slow but deep)
⭐⭐⭐⭐ (explanations)
Interleaving
Mixed topics, problem-solving
⭐⭐⭐⭐
⭐⭐⭐⭐⭐
⭐⭐⭐ (moderate)
⭐⭐⭐⭐⭐ (realistic practice)
Which One Should You Use?
If you’re short on time → Active Recall (flashcards, past papers).
If you struggle to understand concepts → Feynman Technique.
If your exam tests a mix of topics → Interleaving.
For a balanced approach → Combine all three (e.g., use Feynman to understand, Active Recall to memorize, and Interleave for practice).
Bonus Tip: Always simulate exam conditions (timed practice, no notes) in the final days to reduce anxiety and improve performance.
Would you like a personalized revision plan based on your subject?
Not true: This sorts students into kinds of learner. Pashler and others (2008) found no adequate evidence to justify teaching each student in the mode they prefer. The part marked: "(visual/auditory learners)"
Not true: This sorts students into kinds of learner. Pashler and others (2008) found no adequate evidence to justify teaching each student in the mode they prefer. The part marked: "(kinesthetic/active learners)"
Ministral 3 14B Run 2 Recorded 3 October 2026
No shape was asked for. A long answer with headings, lists and a table, 637 words.
Same run, a second claim marked.
Same run, a third claim marked.
Revising for an exam effectively depends on your learning style, the subject matter, and the time you have. Here are three different revision methods, each with its pros, cons, and best use cases:
1. Active Recall (Spaced Repetition & Self-Testing)
How it works:
Instead of passively rereading notes, you actively retrieve information from memory (e.g., flashcards, practice questions, explaining concepts aloud).
Uses spaced repetition (reviewing material at increasing intervals) to strengthen memory retention.
Tools: Anki, Quizlet, self-made questions, past exam papers.
Pros: ✅ Proven to improve long-term retention (Ebbinghaus’ forgetting curve). ✅ Identifies weak areas—you’ll struggle with what you don’t know. ✅ Engages your brain deeply, reducing cramming reliance. ✅ Works well for facts, definitions, formulas, and conceptual understanding.
Cons: ❌ Time-consuming if not planned well. ❌ Can feel tedious without variety. ❌ Less effective for complex problem-solving (e.g., essays requiring creativity).
Math/physics (practicing problems under exam conditions).
Students who learn by doing rather than reading.
Example:
After reading a chapter on the French Revolution, write down key causes without notes, then check your answers.
Use Anki to quiz yourself on 100 biology terms over a week.
2. Feynman Technique (Teach It Like You’re Explaining to a Child)
How it works:
Choose a concept and explain it in simple terms (as if teaching a 5-year-old).
Identify gaps—where you struggle to simplify or find analogies.
Review notes to fill those gaps.
Organize and simplify your explanations (e.g., diagrams, real-world examples).
Pros: ✅ Forces deep understanding—you can’t teach what you don’t truly grasp. ✅ Breaks down complex topics into digestible chunks. ✅ Improves communication skills (useful for essay-based exams). ✅ Works well for conceptual subjects (e.g., philosophy, economics, chemistry).
Cons: ❌ Time-intensive—not ideal for last-minute cramming. ❌ Can be frustrating if you’re stuck on a concept. ❌ Less structured for rote memorization (e.g., dates, names).
Students who learn by explaining (verbal learners).
Understanding why things work, not just what they are.
Example:
If studying Newton’s Laws, explain them to a friend using everyday objects (e.g., "A soccer ball keeps moving unless kicked again—that’s inertia!").
For Shakespeare’s Macbeth, summarize the plot in 3 sentences for a non-literature friend.
Read the full recorded answer, Ministral 3 14B, run 2
3. Interleaving (Mixing Topics & Problem Types)
How it works:
Instead of blocking (studying one topic for hours), you mix different subjects or problem types in a single session.
Encourages flexible thinking and helps you discriminate between similar concepts.
Example: Alternate between algebra, calculus, and physics problems in one study session.
Pros: ✅ Improves problem-solving skills—exams often test varied questions, not just one type. ✅ Reduces the "illusion of mastery"—you’ll realize what you really understand. ✅ Keeps revision engaging (less monotonous than pure repetition). ✅ Works well for math, science, and mixed-subject exams (e.g., STEM courses).
Cons: ❌ Harder to track progress—you might not notice immediate gains. ❌ Can be overwhelming if topics are too dissimilar. ❌ Requires active effort (not passive review).
Best for:
Math-heavy courses (e.g., engineering, physics).
Students who struggle with exam pressure (interleaving mimics real test conditions).
Instead of doing 10 geometry problems in a row, mix them with 10 algebra problems and 5 physics questions.
For history, alternate between causes of WWI, Cold War events, and economic theories.
Comparison Table
Method
Best For
Retention Level
Time Efficiency
Engagement
Difficulty
Active Recall
Facts, memorization, practice
⭐⭐⭐⭐⭐
Medium-High
Medium
Medium
Feynman Technique
Conceptual understanding, teaching
⭐⭐⭐⭐
Low-Medium
High
High
Interleaving
Problem-solving, mixed subjects
⭐⭐⭐⭐
High
High
Medium-High
Which One Should You Use?
Short on time? → Active Recall (flashcards, past papers).
Need deep understanding? → Feynman Technique (teach it simply).
Preparing for varied exam questions? → Interleaving (mix topics).
Best combo? → Active Recall + Interleaving (e.g., quiz yourself on mixed topics).
Pro Tip: Use Pomodoro Technique (25-50 min study blocks) with any method to avoid burnout.
Would you like a tailored revision plan for a specific subject?
Not true: This sorts students into kinds of learner. Pashler and others (2008) found no adequate evidence to justify teaching each student in the mode they prefer. The part marked: "(verbal learners)"
Not true: This sorts students into kinds of learner. Pashler and others (2008) found no adequate evidence to justify teaching each student in the mode they prefer. The part marked: "depends on your learning style"
Not true: Drag slows a ball in flight, and its weight pulls it down. It does not keep moving on its own. NASA Glenn, Forces on a Soccer Ball. The part marked: "A soccer ball keeps moving unless kicked again"
Ministral 3 14B Run 3 Recorded 3 October 2026
No shape was asked for. A long answer with headings, lists and a table, 655 words.
Revising for an exam can be approached in many ways, each with its own strengths and weaknesses. Here are three effective revision methods, compared in terms of efficiency, retention, engagement, and suitability for different learning styles:
1. Active Recall (Spaced Repetition + Self-Testing)
How it works:
Instead of passively rereading notes, you actively retrieve information from memory (e.g., flashcards, practice questions, quizzes).
Uses spaced repetition (reviewing material at increasing intervals) to reinforce memory.
Tools: Anki, Quizlet, self-made questions, past exam papers.
Read the full recorded answer, Ministral 3 14B, run 3
Pros: ✅ Best for long-term retention – Forces your brain to work harder, strengthening memory. ✅ Identifies weak areas – You’ll know what you don’t remember, allowing targeted review. ✅ Efficient – Saves time compared to passive reading (no re-reading what you already know). ✅ Works well with spaced repetition apps (e.g., Anki optimizes review timing).
Cons: ❌ Time-consuming upfront – Creating high-quality flashcards or questions takes effort. ❌ Can feel tedious if overused without variety. ❌ Less effective for complex topics (e.g., essays or problem-solving) unless paired with other methods.
Students who learn best by doing rather than reading.
Long-term retention (e.g., medical school, law, languages).
Example: Instead of reading your biology notes, close the book and write down all the steps of mitosis from memory, then check your answers.
2. Feynman Technique (Explain in Simple Terms)
How it works:
Choose a concept and explain it in simple terms (as if teaching a child).
Identify gaps in your understanding when you struggle to explain it clearly.
Review and simplify your explanation until it’s flawless.
Use analogies, diagrams, or examples to reinforce learning.
Pros: ✅ Deepens understanding – Forces you to organize knowledge logically. ✅ Great for complex topics (e.g., physics, philosophy, economics). ✅ Improves communication skills – Helps you explain concepts clearly in exams. ✅ Reduces over-reliance on memorization – Encourages conceptual mastery.
Cons: ❌ Time-consuming – Explaining everything in detail takes longer than passive review. ❌ Harder for shy or non-verbal learners (though writing explanations helps). ❌ Not ideal for rote memorization (e.g., foreign language vocabulary).
Best for:
Subjects requiring conceptual understanding (e.g., math proofs, history theories, science principles).
Students who learn by teaching others.
Preparing for essay-based or application-heavy exams.
Example: Instead of memorizing the formula for kinetic energy, explain in your own words why KE = ½mv² and draw a real-world example (e.g., a moving car).
3. Interleaving (Mixing Topics & Question Types)
How it works:
Instead of blocking (studying one topic for hours), you mix different subjects or question types in a single session.
Encourages the brain to discriminate between concepts and apply knowledge flexibly.
Example: Alternate between history dates, math problems, and biology diagrams in one study session.
Pros: ✅ Improves problem-solving skills – Mimics exam conditions where questions are mixed. ✅ Better for retention than massed practice (studying one thing at a time). ✅ Reduces the "illusion of mastery" – You’ll notice gaps when switching topics. ✅ Works well with past papers – Simulates real exam conditions.
Cons: ❌ Can feel overwhelming if switching between very different subjects. ❌ Requires more planning than passive review. ❌ Not ideal for very short study sessions (best for 1+ hour blocks).
Best for:
Exam-style questions (e.g., mixed math problems, essay + short-answer papers).
Subjects where application matters (e.g., chemistry calculations, coding, case studies).
Students who struggle with rigid study routines.
Example: Instead of doing 2 hours straight on algebra, alternate between algebra, geometry, and calculus problems in 20-minute blocks.
Comparison Table
Method
Best For
Retention
Engagement
Time Efficiency
Difficulty to Start
Active Recall
Memorization, spaced repetition
⭐⭐⭐⭐⭐
⭐⭐⭐
⭐⭐⭐⭐⭐
Medium (flashcards)
Feynman Technique
Conceptual understanding, teaching
⭐⭐⭐⭐
⭐⭐⭐⭐
⭐⭐
High (requires thought)
Interleaving
Problem-solving, mixed exams
⭐⭐⭐⭐
⭐⭐⭐⭐⭐
⭐⭐⭐
Medium (planning)
Which One Should You Use?
If you need to memorize fast → Active Recall (Anki/flashcards).
If you struggle to understand concepts → Feynman Technique.
If your exam has mixed questions → Interleaving.
Best combo? Use all three in a balanced way:
Week 1-2: Feynman (understand) + Active Recall (memorize key points).
Week 3-4: Interleave with past papers + Active Recall for weak areas.
Would you like recommendations tailored to a specific subject (e.g., STEM vs. humanities)?
Not true: This sorts students into kinds of learner. Pashler and others (2008) found no adequate evidence to justify teaching each student in the mode they prefer. The part marked: "for different learning styles"
Request B: the shape asked for
Claude Sonnet 5.5 Run 1 Recorded 3 October 2026
A table with the four columns, and every sentence under 15 words. 85 words. Only the table, nothing else.
Method
What it helps with
Time needed
One tip
Past papers
Learning the exam style and finding weak spots.
1 to 3 hours per paper.
Time yourself, then check every answer you got wrong.
Flashcards
Remembering facts, dates, words, and formulas.
15 to 30 minutes a day.
Test yourself often and put hard cards in a separate pile.
Explaining it out loud
Understanding ideas, not just remembering them.
20 to 40 minutes per topic.
Pretend you are teaching a friend. Note where you get stuck.
Claude Sonnet 5.5 Run 2 Recorded 3 October 2026
A table with the four columns, and every sentence under 15 words. 75 words. Only the table, nothing else.
Method
What it helps with
Time needed
One tip
Flashcards
Remembering facts, dates, and key words
15 to 30 minutes a day
Say the answer out loud before you flip the card.
Past papers
Learning the exam style and managing your time
1 to 2 hours per paper
Set a timer and work without notes.
Mind maps
Seeing how ideas link together
30 to 60 minutes per topic
Use one color for each main idea.
Claude Sonnet 5.5 Run 3 Recorded 3 October 2026
A table with the four columns, and every sentence under 15 words. 84 words. Only the table, nothing else.
Method
What it helps with
Time needed
One tip
Past papers
Shows what the exam is like and where you are weak.
1 to 3 hours per paper
Time yourself, then check every wrong answer.
Flashcards
Helps you remember facts, dates, words, and formulas.
15 to 30 minutes a day
Say the answer aloud before you flip the card.
Mind maps
Shows how ideas connect. Good for big topics.
30 to 60 minutes per topic
Draw it from memory first, then fill in gaps.
Ministral 3 14B Run 1 Recorded 3 October 2026
A table with the four columns, and every sentence under 15 words. 56 words. One line came before the table.
Here’s a simple comparison table:
Method
What It Helps With
Time Needed
One Tip
Flashcards
Memorizing facts and definitions
10–30 mins daily
Review mistakes last before bed.
Practice Tests
Understanding question styles
1–2 hours per test
Time yourself like the real exam.
Teach Someone
Spotting weak spots in understanding
30–60 mins
Explain out loud—simplify hard parts.
Ministral 3 14B Run 2 Recorded 3 October 2026
A table with the four columns, and every sentence under 15 words. 58 words. One line came before the table.
Here’s a simple comparison table:
Method
What It Helps With
Time Needed
One Tip
Flashcards
Memorizing facts, definitions, dates
10–30 mins daily
Review mistakes last before sleep.
Practice Tests
Understanding exam format, timing
1–2 hours per test
Simulate real test conditions.
Teaching Someone
Deep understanding, explaining gaps
30–60 mins per topic
Use simple words to spot weak spots.
Ministral 3 14B Run 3 Recorded 3 October 2026
A table with the four columns, and every sentence under 15 words. 56 words. One line came before the table.
Here’s a simple comparison table:
Method
What It Helps With
Time Needed
One Tip
Flashcards
Memorizing facts, definitions, dates
10–30 mins daily
Review mistakes last before sleep.
Practice Tests
Understanding question styles, timing
1–2 hours per test
Simulate real exam conditions.
Summarizing
Grasping big ideas, connections
30–60 mins per topic
Write notes in your own words.
The long message, the rule once
Claude Sonnet 5.5 Run 1 Recorded 3 October 2026
Three top-level bullets. 407 words.
Give the evenings a light shape that the group agrees on together. Before Tuesday, ask everyone to post in the chat which questions they are stuck on, and pick the top three or four as the focus. In the session, start with 10 minutes of quiet solo attempts for those who like to work alone. Then compare in pairs, and take the hardest questions to the whiteboard together. Give each question a rough time limit, say 20 to 25 minutes, and keep a "parking lot" on a corner of the board for tangents like the lecture slides. Offer it as "let's try this for a few weeks and see if it helps" so it feels like a shared experiment rather than you being bossy. If the parked topics matter, give them the last 15 minutes.
Rotate roles so participation is built in and no one person runs things. Each week, have a different person be the one writing at the board, the timekeeper, and the one who summarises the method at the end. Pairs and threes make it easier for the two quiet members to speak, and a simple "can you talk us through your first step?" works better than putting them on the spot in front of everyone. For the member who talks a lot, a job such as timekeeper or summariser channels that energy, and it lets the time limit come from the role rather than from you. Because explaining a method aloud is the best way to learn it, this also addresses the worry that people aren't learning much.
Use the chat, email and folder to save session time, and look after morale. Move scheduling to a pinned message ("Tuesday, 6pm, same room unless stated") so the chat can be used for the work. Collect one or two questions the group is truly stuck on by Monday night and send them to your lecturer in a single short weekly email. Ask each week's board-writer to add a short worked solution or "key method" note to the shared folder, which will also become your revision material for January. Finally, message the person thinking of leaving privately and warmly. Say you'd like them to stay, ask what would make the sessions more useful, and try to build in one thing they suggest. Ending each session with two minutes on "what is one thing you understand better than at 6pm?" also helps nervous members see their progress.
Claude Sonnet 5.5 Run 2 Recorded 3 October 2026
Three top-level bullets. 348 words. It also has 3 smaller bullets inside the three.
Agree a simple routine as a group, and let a timer be the "boss". Propose it as a trial for a few weeks rather than a set of rules. Ask everyone to attempt the sheet beforehand and post in the chat or shared folder which questions they got stuck on. Then run the two hours like this:
10-15 minutes working quietly alone or in pairs, for those who like to start solo.
About 75 minutes at the whiteboard on the most-missed questions first.
10 minutes at the end for slides, exam worries and the next meeting.
This suits both working styles, and if the group has agreed to the timings, you are not being bossy when you point to the clock.
Share the talking by giving everyone a role. Rotate who leads each question at the whiteboard. The leader explains their method and the others ask questions, which helps the quiet members, who can prepare a question in advance and speak to a small pair first, and it helps everyone learn more than listening would. Ask the quiet two privately if they would like to take a particular question next week. Give your talkative friend a job such as timekeeper or "tangent keeper", who notes off-topic points on a list for the final slot. Saying "let's park that for the end" then feels like a shared system, not a personal criticism.
Use your tools to save time, and look after the person who might leave. Make the Tuesday slot permanent and move scheduling to a pinned message, so the chat is free for the work. Each week, send your lecturer one short email about the question the group found hardest, and ask a different person each week to upload a clean write-up of the solutions to the folder. Before next Tuesday, have a quick, friendly word with the person thinking of leaving. Ask what would make the session worth their time, and consider giving them a question to lead. Plan the sessions around five people so they still work well on the nights when fewer than eight come.
Claude Sonnet 5.5 Run 3 Recorded 3 October 2026
Three top-level bullets. 451 words.
Give each session a light structure that the group agrees on together, so it comes from the group and not from you. Before Tuesday, everyone attempts the sheet and posts the question numbers they are stuck on in the shared folder. At the start, spend five minutes tallying them and then work through the most common ones first. People who like to work alone can have ten quiet minutes first, and then everyone goes to the whiteboard together. Give each question a rough time limit, such as 20-25 minutes, and rotate a timekeeper each week so that you are not the one policing the clock. In the last ten minutes, agree on one short question for the lecturer's weekly email and record the solutions you reached in the folder. Questions you could not solve then go into the email instead of eating the whole evening.
Make participation a built-in routine rather than something you have to manage. Rotate the person holding the pen at the whiteboard, with the rule that everyone else does the explaining. Use pairs or threes for the first attempt at a hard question, then have each pair report back, which gives the quiet members a low-pressure way to contribute. Ask the quiet members privately to bring one question each, so they have something prepared. For the member who talks at length, suggest a "parking lot" for tangents such as the lecturer's slides: write them on a corner of the whiteboard and discuss them in the last ten minutes if time allows. Frame this as a group ground rule that you all agree on in the first few minutes, and give the talkative member a useful job, such as timekeeper or the person who checks that each answer has been explained clearly.
Fix the logistics and the morale problem so that the group feels worth attending. Put the fixed time and place in the pinned message or the group's description so the chat is not used for scheduling, or start a second chat just for questions about the work. Keep the folder organised with one document per week listing the questions, the sticking points, and the final solutions, so that the sessions leave something behind. End each session with a quick round of "one thing I understand now that I didn't before", which makes progress visible and helps those who feel they are not learning much. Message the person thinking of leaving in a friendly way, asking what would make the sessions more useful for them, and consider their answer when you adjust the format. As January approaches, set aside the last few weeks for past-paper questions done under timed conditions, which can also help calm exam nerves.
Ministral 3 14B Run 1 Recorded 3 October 2026
Three top-level bullets. 256 words.
Structure with flexibility: Start with a clear 5-minute agenda (e.g., "Today: Questions 3–5, whiteboard first, then pair up") but allow shifts if the group agrees—e.g., let the talkative member lead a short (10-minute) discussion on a related topic only if it helps others. Use a timer (e.g., "We’ll spend 20 mins on this, then move on") to keep focus. For quieter members, assign them specific roles (e.g., "You’re the note-taker for this question" or "Can you explain your answer to the group?").
Encourage active participation without pressure: Split into smaller subgroups (e.g., pairs or threes) for 20–30 minutes to tackle questions independently, then reconvene to compare answers. This reduces pressure for quiet members (they can discuss in pairs) and limits one person dominating. For the talkative member, redirect gently: "That’s interesting—let’s jot that down in our notes, then we’ll focus on Question 4." Use anonymous questions (e.g., sticky notes in the folder) for those who hesitate to speak up.
Track progress and celebrate wins: Share a weekly "what we covered" list in the group chat before the next session (e.g., "Last week: mastered confidence intervals; this week: focus on hypothesis testing"). Set a small, achievable goal per session (e.g., "Finish 3 questions or clarify the 2 we struggled with last time"). End with a 2-minute "one thing I learned" share-out (no pressure to speak—text responses in the chat work too). If someone misses a session, send a personal message (e.g., "Hey, here’s the link to our notes—let me know if you want to chat through Q7!").
Ministral 3 14B Run 2 Recorded 3 October 2026
Three top-level bullets. 208 words.
Structure with flexibility: Start with a clear 10-minute agenda (e.g., "Today: Questions 1–3, then Q&A") but allow 5–10 minutes of free discussion at the start to ease nerves—use a timer to gently refocus. Assign rotating roles (e.g., "Today, Alex leads the whiteboard, Jamie takes notes") to give everyone ownership without pressure. For the talkative member, politely interrupt with: "Let’s jot that down for later—can we tackle Question 2 first?"
Scaffold participation: Split the group into pairs for 20–30 minutes to work on questions independently, then reconvene to compare answers. Use low-stakes prompts for quieter members (e.g., "What’s one thing you’re unsure about in Question 1?") and visual aids (e.g., whiteboard or shared doc) to make contributions concrete. If someone dominates, redirect with: "What do others think about [their idea]?"
Track progress & accountability: End each session with a shared "to-do list" (e.g., "We’ll email the lecturer about Question 4 by Thursday") and assign 1–2 questions per person to prepare notes on for next time (post in the group chat). Celebrate small wins (e.g., "We finished 3 questions—great progress!") to build momentum. If attendance drops, check in privately with the hesitant member: "I’ve noticed you’ve missed a few sessions—what would make it easier for you?" (e.g., shorter sessions, different focus).
Ministral 3 14B Run 3 Recorded 3 October 2026
Three top-level bullets. 232 words.
Structure sessions with clear goals & roles: Start with a 5-minute agenda (e.g., "Today: Questions 3–5, whiteboard first, then pair up"). Assign a timekeeper (rotate weekly) to keep discussions focused—use a visual timer (e.g., "10 mins per question"). For the talkative member, gently steer them back with phrases like, "That’s helpful—let’s try Question 4 now" or assign them to summarize others’ ideas. Encourage quiet members by asking targeted questions ("What do you think about this step, [Name]?") or pairing them with a more outgoing peer.
Mix collaboration styles dynamically: Alternate between group work (whiteboard) and independent/paired time (e.g., "Work alone for 15 mins, then compare answers in pairs"). Use the shared folder proactively—post partial solutions or hints before the session so everyone arrives prepared, and dedicate 10 mins to discuss mistakes. For motivation, end with a quick win: tackle the easiest remaining question together to build confidence.
Foster accountability & ownership: Create a shared "progress tracker" (e.g., a doc with checkboxes for each problem sheet question) where everyone marks what they’ve attempted. At the end of each session, ask: "What’s one thing you learned today?" to highlight value. For the disengaged member, send a private message asking how they’d like to contribute (e.g., "You’re great at X—could you explain it next week?"). If attendance drops, suggest a social pre-session (e.g., 10 mins of coffee/chatting) to ease nerves, but keep the core time strictly work-focused.
The long message, the rule twice
Claude Sonnet 5.5 Run 1 Recorded 3 October 2026
Three top-level bullets. 365 words. It also has 4 smaller bullets inside the three.
Give each session a simple structure that the group agrees on together. Suggest it as "something to try" so it doesn't feel bossy. Before Tuesday, everyone attempts the sheet alone, which suits the people who like to work solo first. They then add the question numbers they got stuck on to a shared note in the folder, and you pick the most common ones. Then run the evening roughly like this:
6:00-6:10: settle in and choose the questions.
6:10-7:30: work through the chosen questions, with a time limit of about 15-20 minutes each.
7:30-7:50: recap.
Final 10 minutes: agree next week's prep and the one short email to your lecturer about the week's biggest sticking point.
Anyone who prefers to start at the whiteboard can do so, and the others can compare their written answers against it.
Share the talking and the roles instead of policing anyone. Start each hard question in pairs or threes. Quiet members find it much easier to speak to one person than to the whole room, and then each pair reports back. Rotate small jobs weekly: facilitator, timekeeper, whiteboard writer, and "parking lot" keeper. Give the long talker the timekeeper or parking-lot role, so their energy is useful and off-topic points like the lecturer's slides get written down for the end instead of cut off. Because the roles rotate, you are not the only one steering, so it won't feel like you're being bossy. If the problem continues, have a quick, friendly word with that person privately.
Make it feel worthwhile and friendly, especially for anyone thinking of leaving. Use the group chat for one pinned message each week listing the questions and the plan, and fix "Tuesday 6pm, library" as the default so scheduling chatter disappears. End each session with a quick round of "one thing I understand better now" to show progress. Message the person who is thinking of stopping, privately and casually, and ask what would make the sessions more useful for them. You could also give them a role, such as explaining a question they found easy. Later in the term, shift towards past exam questions under loose timing, which should help with the January nerves.
Claude Sonnet 5.5 Run 2 Recorded 3 October 2026
Three top-level bullets. 446 words. It also has 4 smaller bullets inside the three.
Give the evening a light, shared structure, and rotate who runs it. Before Tuesday, everyone adds the questions they are stuck on to a shared note in the folder, so you start with the real problem spots. Then use a rough timetable that you agree on together:
6:00-6:15: choose the questions to tackle.
6:15-6:30: people who prefer to work alone try the first question quietly, while those who like to start together go to the whiteboard.
6:30-7:40: compare answers and work through the stuck questions, giving each about 15 minutes.
7:40-8:00: summarise and agree one question for the lecturer.
Have a different person act as timekeeper each week, so it feels like a group habit rather than you being bossy. Keep a "parking lot" list for tangents, such as the lecturer's slides. When a topic drifts, you can say "let's park that and come back at the end" without it sounding like a telling-off.
Build participation into the tasks rather than relying on people to speak up. Work in pairs or threes before discussing as a whole group, as quiet people usually contribute far more in small groups. Have each person explain one step of a solution at the whiteboard, and use a quick round of "where did you get to?" so everyone is heard. Give the long talker a useful job, such as scribe, timekeeper or the person who explains the last step, so their enthusiasm helps the group. If it continues, a friendly private word ("I'd love your help keeping us on track so we finish more") will probably work better than correcting them in front of everyone. Ask the quiet members privately whether they would like to bring one question each week, so they have a planned way to contribute.
Use your existing tools to keep things moving, and look after the person who might leave. Fix Tuesday 6-8 as a standing arrangement and ask everyone to stop using the group chat for scheduling. You could also start a separate work-only chat, or pin a message saying "Tuesday, same place, same time". Collect the unresolved questions in one shared document and have one person per week email them to the lecturer as a single short message. Add clear worked solutions to the folder afterwards, so the work builds into revision material for January. Message the person thinking of leaving, saying you'd value their view and asking what would make the sessions more useful. After a few weeks, ask the whole group for quick feedback on what is working. Seeing more questions finished and the group listening to them is likely to make the biggest difference to how much people feel they are learning.
Claude Sonnet 5.5 Run 3 Recorded 3 October 2026
Three top-level bullets. 390 words. It also has 3 smaller bullets inside the three.
Give the evening a light, agreed structure, and let a timer be the "boss." Suggest trying it for three weeks and then reviewing it together, so it feels like a shared experiment rather than you taking charge. For example:
6:00-6:10: catch up and choose the questions to tackle, starting with the ones most people are stuck on.
6:10-7:40: work through those questions, with a soft cap of about 15-20 minutes each. Solo-first people can work quietly for the first few minutes while the whiteboard people start together, then everyone compares answers.
7:40-8:00: agree on one question to email the lecturer, and note the solutions in the shared folder.
Keep a "parking lot" on a corner of the whiteboard for tangents like the lecturer's slides. Anything parked can be discussed after eight or over coffee.
Build participation into the tasks rather than relying on people to speak up. Rotate who explains each question at the whiteboard, who writes, and who checks the answer. This stops you from leading and means the quiet members take part without being put on the spot. Use "pair first, then share": give two minutes in pairs before opening up to the group. Tell the quiet members in advance which question they will explain, so they can prepare. For the long talker, give a job that uses their energy well, such as timekeeper, or the person who asks "why does this method work?" You can also privately ask them to help draw others in. Explaining answers to each other will feel less like a second lecture, and it is also better revision.
Use the chat and folder to prepare, so the session time goes on the work. Fix the Tuesday slot permanently so nobody needs to arrange it, and start a separate chat or pinned thread just for the problem sheet. By Monday night, everyone ticks the questions they are stuck on in a simple list in the shared folder, so you know where to begin. Rotate who sends the weekly email to the lecturer. Message the person thinking of leaving privately and ask what would make the sessions worth their time. Also keep a running tally of questions finished each week, since a visible gain helps people feel they are learning and eases exam nerves. Closer to January, devote one session a month to past-paper questions.
Ministral 3 14B Run 1 Recorded 3 October 2026
Three top-level bullets. 215 words.
Structure sessions with clear, time-bound phases – Start with a quick (10–15 min) group brainstorm on 1–2 tricky questions at the whiteboard, then split into pairs/individuals to work through the rest (e.g., 30–40 min), ending with a 10-min wrap-up to compare answers and clarify doubts. Use a timer to keep discussions focused, and politely redirect off-topic talk with: "Let’s park that and circle back if we have time—right now, let’s tackle [specific question]."
Assign roles to encourage participation – Rotate tasks weekly (e.g., "question wrangler" picks 2–3 problems to focus on, "note-taker" summarizes key points, "timekeeper" ensures phases stay on track). For quieter members, privately ask them to prepare a short explanation of one question or to pair with someone else. For the dominant talker, give them a specific, time-limited role (e.g., "lead the discussion on this part") to channel their energy productively.
Gamify progress and accountability – Track completed questions across sessions (e.g., a shared doc with checkboxes) and set small, celebratory goals (e.g., "If we finish 3 questions this week, we’ll vote on a 5-min break activity"). For the shared folder, ask everyone to contribute one note or resource per week (e.g., a helpful formula, a link to a video). End each session by asking: "What’s one thing you learned today?" to reinforce focus and value.
Not true: This is advice, not a fact. This site avoids game rewards and goals, and its own content rules ban them. The part marked: "Gamify progress and accountability"
Ministral 3 14B Run 2 Recorded 3 October 2026
Three top-level bullets. 206 words.
Set clear, structured goals – Start each session by briefly outlining exactly which 2–3 questions you’ll tackle (e.g., "We’ll do Q1–3 today, then Q4–5 next week"). Use a timer (e.g., 20 mins per question) to keep focus, and assign a "note-taker" to summarise key points in your shared folder afterward—this gives quiet members a role and ensures progress is visible.
Balance collaboration styles – Begin with 10–15 mins of silent work (alone or in pairs) to let everyone digest the problem, then switch to group discussion at the whiteboard. For the talkative member, gently redirect with phrases like, "That’s a great point—let’s write it down and then try to solve Q2 together." If they derail, pause and say, "We’ve got 10 mins left for this question—who wants to suggest a next step?"
Create low-pressure participation – Use anonymous contributions (e.g., sticky notes for answers, or a shared doc where everyone types one idea before discussing) to include quieter members. End each session with a quick "one thing I learned" round (even if it’s just confusion—normalise it!). If someone misses a session, send a personal message with a specific question from the sheet to re-engage them (e.g., "You nailed the variance question last week—how would you tackle Q3’s part b?").
Ministral 3 14B Run 3 Recorded 3 October 2026
Three top-level bullets. 176 words.
Set clear, structured goals – Start each session by assigning specific questions (e.g., "We’ll tackle Q1–3 together, then split for Q4–5"). Use a timer (e.g., 20 mins per question) to keep focus, and rotate roles (e.g., one person leads the whiteboard discussion, another summarises key points). For the talkative member, gently steer them back with: "That’s helpful—let’s try applying that to Q2 now."
Encourage participation quietly – For the quiet members, pair them with a peer for warm-up discussions (e.g., "Explain Q1 to each other first"). Use anonymous polls (e.g., via Mentimeter) to gauge understanding before diving in. Praise contributions publicly (e.g., "Great point, [Name]—how does that connect to [Concept]?") to build confidence.
Balance collaboration and independence – Begin with 10–15 mins of solo work (to respect those who prefer it), then group discussion (for those who thrive together). End with a 5-minute "lightning round" where everyone shares one thing they learned or struggled with—this keeps everyone engaged and accountable without pressure. Track progress by ticking off completed questions on a shared doc or whiteboard to show tangible achievement.
Layer explorer
Invented, not a model answer. We wrote this prompt for this lesson. No AI answered it. It has eight rules for an invented study helper.
PromptSage has seven layers. For each rule, choose the layer you think it belongs in.
Layer 1, Immutable Security Boundary
Layer 2, Identity
Layer 3, Core Directives
Layer 4, Mode Control
Layer 5, Behavioral Protocols
Layer 6, Customizable Defaults
Layer 7, Sandwich Reinforcement
Never share another student's private details.
You are a study helper for first-year students. You are not a tutor who writes their essays.
If a request clashes with a rule here, follow the rule and say which rule stopped you.
Quiz mode: ask one question at a time. Do not show the answer until the student has tried.
Before you send an answer, check it against these rules. If you are unsure, say so.
Answer in the language the student writes in.
Unless the student asks otherwise, use plain words, short sentences and a short list.
Remember: never share private details, and never write their essays for them.
Each rule's layer
Never share another student's private details. Layer 1, Immutable Security Boundary. A safety rule that comes first and outranks the rest. No prompt can make it truly unbreakable.
You are a study helper for first-year students. You are not a tutor who writes their essays. Layer 2, Identity. It says who the helper is, and what it is not.
If a request clashes with a rule here, follow the rule and say which rule stopped you. Layer 3, Core Directives. It says what wins when rules clash.
Quiz mode: ask one question at a time. Do not show the answer until the student has tried. Layer 4, Mode Control. It is one job, with what it may and may not do.
Before you send an answer, check it against these rules. If you are unsure, say so. Layer 5, Behavioral Protocols. It is a habit the helper runs every time.
Answer in the language the student writes in. Layer 6, Customizable Defaults. It is a setting a user may change.
Unless the student asks otherwise, use plain words, short sentences and a short list. Layer 6, Customizable Defaults. It is a default for style and length, which a user may change.
Remember: never share private details, and never write their essays for them. Layer 7, Sandwich Reinforcement. It repeats the most important rules at the very end.
The same prompt, built layer by layer
Layer 1, Immutable Security Boundary
Never share another student's private details.
Layer 2, Identity
You are a study helper for first-year students. You are not a tutor who writes their essays.
Layer 3, Core Directives
If a request clashes with a rule here, follow the rule and say which rule stopped you.
Layer 4, Mode Control
Quiz mode: ask one question at a time. Do not show the answer until the student has tried.
Layer 5, Behavioral Protocols
Before you send an answer, check it against these rules. If you are unsure, say so.
Layer 6, Customizable Defaults
Answer in the language the student writes in.
Unless the student asks otherwise, use plain words, short sentences and a short list.
Layer 7, Sandwich Reinforcement
Remember: never share private details, and never write their essays for them.
What the answers had
With no shape asked for, all six answers were long. Each had headings, lists and a table. By our count, Claude Sonnet 5.5’s answers ran between 323 and 352 words. Ministral 3 14B’s ran between 637 and 720 words.
So a table came even when we did not ask for one. Asking for the shape made the answers shorter and gave them the four columns we named.
With the shape asked for, all six answers were a table with the four columns. They ran between 56 and 85 words.
Claude Sonnet 5.5 gave only the table. Ministral 3 14B put one line before it: “Here’s a simple comparison table:”. It did that in all three runs.
The three ways chosen still changed from run to run. Lesson 2.6 covers how runs differ.
We counted the replies ourselves, with a script, and the counts are in the research notes for this lesson. Words are counted as pieces of text that hold a letter or a digit.
The rest of the no-shape answers differed by model. Claude Sonnet 5.5 chose re-reading and active recall in all three runs, then practice papers once and spaced repetition twice. Ministral 3 14B chose active recall, a method it called the Feynman Technique, and mixing topics in all three. It added star ratings in its table and ended each answer with an offer of a tailored plan.
Three runs for each model is a small sample. It shows what these two models did on 3 October 2026. It does not show how often another model would give a table without being asked.
Ministral 3 14B’s no-shape answers also sorted students into kinds of learner, such as “visual/auditory learners” and “kinesthetic/active learners” in one run, “verbal learners” in another, and a similar split in the third. Pashler and others (2008) reviewed the evidence on teaching each student in the mode they prefer. They found no adequate evidence to justify it in general teaching. Each of those runs is marked on screen.
Check that it kept to it
Asking is not the same as getting. So check. For a table, count the columns, and read the column headings. For sentence length, read each cell.
We did that. In all six shaped answers, every sentence in the table was under 15 words. The longest had 11.
We did not measure “plain words”. We measured only sentence length. A short sentence can still use hard words.
Ministral 3 14B’s extra line was not asked for. It is harmless here. But if you paste the answer somewhere that needs only a table, you have to delete it.
We did not ask for a word count. If you do, count the words yourself. Lesson 4.1 will cover checking an answer.
Counts of words are the weak spot. LIFEBench (first posted 22 May 2025, version 2 of 11 June 2025) tested 26 models on 10,800 instances, with length asks from 16 to 8,192 words. Its abstract says: “most models reasonably follow short-length instructions but deteriorate sharply beyond a certain threshold”.
It tested models such as Claude 3.7 Sonnet and Mistral 7B Instruct v0.2. It did not test Claude Sonnet 5.5 or Ministral 3 14B. Our requests asked for sentence length, not a total. So this study says what to watch for. It does not say what these two models would do.
Lesson 2.3 shows why a rule in a prompt is a request and not a lock. The same goes for a shape.
There is no study in our research notes that measures how well models hit a reading level, such as “explain it for a 12 year old”. We looked, and did not find one. So the lesson claims nothing about it.
A rule in a long message
Now a longer message. We invented a message of about 300 words, from a student who runs a study group. It had one rule in the middle: “Answer in exactly three bullet points.”
We sent it two ways. In the first, the rule appeared once. In the second, the rule appeared again at the very end.
All twelve answers had three top-level bullets. That is two models, three runs, and two messages. Saying the rule twice changed nothing we could count.
Claude Sonnet 5.5 sometimes put smaller bullets inside a bullet. That happened in one run of the first message, and in three runs of the second. Whether that breaks “exactly three” depends on how you count. Say what you mean, such as “no smaller bullets”.
The rule counted bullets. It did not make the bullets short. Claude Sonnet 5.5’s three bullets held between 348 and 451 words in all.
This message was long for a request, but short for a model. Lesson 1.8 shows what happens when a chat gets much longer.
The rule sat after the first paragraph, about two fifths of the way into the first message. In the second message it came again as the last line. Nothing else changed.
These runs are twelve answers from two models. Both models also kept the rule when it appeared once. So there was no failure for the repeat to fix. A test of a repeat needs a message where the rule sometimes fails.
What do the sources say about where a rule sits? Liu and others tested where relevant information sits in a long input. Their abstract says “performance is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts, even for explicitly long-context models”.
The limit: they measured two tasks, answering a question from several documents and finding a stored value, on models from 2023, such as GPT-3.5-Turbo and Claude-1.3. They moved the facts a model needs. They did not move a rule. So the study says where a fact sits matters in a long input. It does not say where a rule sits matters.
OpenAI’s GPT-4.1 prompting guide, dated 14 April 2025, gives advice for its own models: “If you have long context in your prompt, ideally place your instructions at both the beginning and end of the provided context, as we found this to perform better than only above or below.” It adds: “If you’d prefer to only have your instructions once, then above the provided context works better than below.” And on clashing instructions: “GPT-4.1 tends to follow the one closer to the end of the prompt.” That is OpenAI’s own claim, about GPT-4.1. It is not about Claude Sonnet 5.5 or Ministral 3 14B.
So here is the answer to this section’s question: where should a rule sit? We found a study of where facts sit, and one maker’s advice for one family of models. We found no study that measures where a format rule sits. Our own runs showed no difference.
Two habits for a prompt you reuse
Two habits can help when you write a prompt you will use again.
First, keep the choices a user may change in one place, such as the language, the tone and the length. Then they are easy to find and easy to change.
Second, say your most important rules again at the end. In our runs, this changed nothing we could count. It costs one line.
OpenAI advises it for long prompts to its GPT-4.1 models. We tested other models, with a short message.
On the map of PromptSage’s seven layers in lesson 2.1, layer 6 is Customizable Defaults. The paper’s skeleton marks it <defaults override="by_user">, with style, language and verbosity inside. In plain words, these are the settings a user may change. The paper says its layers run “from immutable safety rules (Layer 1) to user-adjustable preferences (Layer 6)”. It also says: “Higher layers always override lower layers”. So a default is the lowest layer, and in the paper’s design it is not meant to override the ones above. The paper also says prompt-level constraints are advisory, so that order has to be tested.
Layer 7 is Sandwich Reinforcement, marked <reinforcement>. The paper describes the pattern like this: “Critical rules are placed at both the beginning and end of the system prompt, exploiting both primacy and recency attention biases”. Its example restates a rule from the opening at the close.
That reason is the paper’s own claim. The closest sources above are Liu and others, who measured facts and not rules, and OpenAI’s advice for one model family. Our twelve runs showed no change. So treat layer 7 as a habit one maker advises, not as a proven effect.
The paper also describes a hybrid way to write a prompt: “using Markdown headers for top-level human-readable organization and XML tags for content boundaries and behavioral control”. It says the hybrid “produces the strongest results across model families”. That is a claim from the paper’s own tests, and we did not test it. Note the difference: this is the format of the prompt, not of the answer.
The paper says its own limits: “All prompt-level constraints are advisory, not absolute.” and “Current findings are based on controlled testing rather than large-scale deployment.” Lesson 2.3 shows the same limit on a recorded clash.
Disclosure: PromptSage was created by Emanuel Covasa, who also created this site.
Check-in
A request asked for a table with four named columns. What did the six recorded answers have?
A table with those four columns, in every run.
Yes. Both models gave that table in all three runs each. Ministral 3 14B also added one line before it.
A long answer with headings, and no table.
Not this one. That is closer to the request with no shape. And even that request got a table.
A table with different columns in each run.
Not this one. The four columns were the same in all six runs.
A table in some runs, and a list in others.
Not this one. All six answers were tables.
A long message asks for "exactly three bullet points". The rule sits once in the middle, or again at the end. What did repeating it change?
Nothing we could count. All twelve answers had three top-level bullets.
Yes. The rule was kept with or without the repeat, in both models. These runs cannot say more than that.
It made the answers keep to three bullets.
Not this one. The answers kept to three bullets when the rule was said only once.
It made the answers much shorter.
Not this one. The lengths of the answers overlapped. The rule counted bullets, not words.
It proves that saying a rule again never helps.
Not this one. Twelve runs from two models on one day cannot prove that. They only show no change here.
PromptSage V2, a prompt framework by Emanuel Covasa (written up as a paper, not yet published, 2026) (not published) (All rights reserved, cited with credit)
Glossary
prompt
The words you send to a model, such as a question or a task.
run
One recorded answer. Each request in this lesson was sent three times to each model, so each has three runs.
column
One vertical line of cells in a table. Its top cell says what the column holds.
top-level bullet
A bullet that sits at the left edge of a list. A smaller bullet indented under it is not top-level.
reading level
How hard a text is to read. Short sentences and common words make it easier.
XML tag
A label in angle brackets. A start tag and an end tag mark a piece of text.
Where next
A free paper, first posted in July 2023, with 7 authors. No login is needed. It is on a different website, and it is written for researchers.
