Module 2 · About 30 minutes

Effective prompt writing

Think not about what AI gives you, but about what you tell it. In this module we first understand how the model sees the world (tokens and context window), then master the single variable that determines output quality: the instruction.

You can also complete this module by listening
The audio summary of this module was produced with the Madlen Podcast tool.
0:00 0:00
The turning point

Same model, same tool, different result.

Two different instructions on the same topic. The difference isn't ability, it's specificity.

Prompt A · bland version
Write a lesson explanation for photosynthesis.

Result: encyclopaedic language, disconnected from the class level, a text with no specified duration or format. Reads like a quote from any textbook.

Prompt B · specific version
Write a Year 8 science photosynthesis lesson in the voice of a teacher walking up to the board. Duration 8 minutes. Start with an everyday example (the bread at breakfast), then the equation, then debunk 2 misconceptions. Sentences should be short and close to spoken language.

Result: an explanation in Year 8 vocabulary, in a spoken rhythm, bridging from concrete to abstract and targeting misconceptions from the outset.

0
Accuracy increase with structured prompts
White et al., 2023
0
Quality jump for beginners using templates
Zamfirescu-Pereira, CHI 2023
0
Gap between novice and expert prompter
Ethan Mollick, Wharton
Technical breath · Interactive

Token: the unit of the model's world.

The model doesn't read sentences the way you do; it breaks them into meaningful fragments. These fragments are called tokens. Click the words to see how they're broken up.

English
photosynthesis
1 word · approximately 3 tokens
Another example
understanding
1 word · approximately 4 tokens

Words are broken into sub-word pieces. The model processes these tokens rather than whole words, which is why longer technical words cost more tokens. Being fluent in your prompts makes them both cheaper and clearer.

Interactive · Split your own text

How many tokens is your sentence?

Those two words were prepared examples. Now write or paste your own prompt: it gets broken into tokens so you can see the cost with your own eyes. The split is approximate, real models break slightly differently, but the order of magnitude is right.

Technical breath 2

Context window: the model's working memory.

The number of tokens the model can hold in mind at the same time. It grew a hundredfold in three years, for a teacher, this means "being able to place all the relevant documents in front of it".

GPT-3.5 2022 · ~3 pages
4,000
GPT-4 2023 · ~96 pages
128,000
Claude 3.5 2024 · ~150 pages
200,000
Claude Opus 4.6 2026 · 500 pages, a novel's worth
1,000,000

Teacher translation: "I'm placing this learning objectives list, this unit plan, last week's test, and 3 book chapters in front of you. Evaluate all of them together and design a lesson for this week."

The model is like a workdesk.

The best analogy you can think of: a teacher's desk. Hover over the cards.

A · Bookshelf (training data)
Everything the model has ever read
Always there. But when it will recall what is unpredictable.
B · Desktop (context window)
The papers currently open
Your prompt, files you've uploaded, chat history. This is where the model actually "looks".
C · Fallen off the desk (limit exceeded)
When the desk is full, the oldest paper falls off
This is why the model "forgets" instructions from the beginning of long conversations.

A large window sees a lot. But keeping the desk tidy is still the teacher's job.

Core model

Five components, one checklist.

You don't always need to use all five. Think of this as a list to come back to when you're stuck.

1
Role
Who are you?
2
Goal
What will the student gain?
3
Context
For whom, under what conditions?
4
Format
What will the output look like?
5
Constraint
What should it not include?
Other recipes for the same job

No single framework, but they all say the same thing.

Madlen's five-component framework isn't the only right answer. There are two other short recipes teachers frequently use around the world. All share the same message: tell the model who you are, what you want, and for whom/how to produce it.

Madlen · 5 components
Role · Goal · Context · Format · Constraint

The list you just saw. Use it like a checklist, you don't always need all five.

POP method
Persona · Objective · Parameters

Persona (role), objective (task), parameters (audience, tone, reading level). Easy to remember in three steps.

Five-part recipe
Role · Purpose · Specificity · Source · Refine

Works well for research tasks: ask for a "Source" (and verify it's real), then "Refine", meaning improve iteratively.

Interactive · Prompt builder

Build your own prompt with five components.

Make a choice for each component; your prompt takes shape live below. When done, copy it, try it in Madlen Assistant, then start adapting it to your own lesson.

  • (leave empty)
  • You are a Year 9 maths teacher.
  • You are a secondary physics teacher who values connecting lessons to everyday life.
  • You are a primary school teacher with students who have additional learning needs.
  • You are a Year 6 teacher preparing students for transition to secondary school.
  • (leave empty)
  • Students should be able to convert between fractions and decimals.
  • Students should be able to explain the three main stages of photosynthesis in their own words.
  • Students should be able to compare two sources and identify different perspectives.
  • (leave empty)
  • Mixed-ability class of 30 Year 7 students, 40-minute lesson.
  • No devices available, only whiteboard, pen and paper.
  • Inclusive classroom with a student who has a hearing impairment, verbal-only delivery.
  • (leave empty)
  • Give the output as a 3-column table: concept, definition, everyday example.
  • Include an answer key in brackets under each question.
  • 5 bullet points, each no longer than 2 sentences.
  • (leave empty)
  • Do not use mathematical symbols; write everything in words.
  • Keep it under 200 words.
  • Use only examples familiar to UK students.
Live prompt
Interactive exercise

Fix the bad prompt.

The most popular part of the workshop. Which pieces of information should be added to the weak prompt below? Select everything that's missing, then check.

Interactive · Weigh your own prompt

How many components does your prompt carry?

Repairing a prepared prompt is easy; auditing your own is hard. Write a prompt you would genuinely use below and it gets scanned against the five components. The scan works at word level, so it is a checklist rather than a grade, but it catches a missing component surprisingly well.

0 / 5 components
Start typing, the scan runs live.
Interactive · Choose your own adventure

Nine prompt tools, three levels.

Nine techniques to strengthen your prompt once you've built the five-component framework. Pick the level you're ready for, copy the real classroom example under each tool, and try it in Madlen Assistant. Start at the beginning, progress as you master each one.

Interactive · Step-by-step iteration

Same task, from mediocre to excellent in three rounds.

The first prompt is rarely the last. Watch how we refine a seating plan over three rounds. Track what gets added to the prompt in each round and how the output improves.

Advanced level

Three more techniques.

1 Teach with examples
Instead of explaining the rule, show an example; the model picks up the pattern and repeats it. Especially effective for format and tone. A single example is far more powerful than a half-page explanation of the rule. "Write 3 more short stories in the style of the following: [example story]" This technique even has a medical equivalent: the PRET system developed by Hong Kong and Harvard classifies 18 different cancers with up to 100% accuracy by showing the model a few tissue images. The result is determined not by the power of the tool, but by how you use it.
2 Ask it to think
Additions like "solve step by step", "summarise the solution method first, then apply it" stop the model from rushing to an answer. One of the most powerful tactics for reducing hallucination when generating maths and science questions. "Solve this problem. First explain the solution method step by step, then give the final answer."
3 Iterative refinement
The first output is a starting point, not a finish. "This is good but paragraph 2 is too abstract, replace it with a concrete example a Year 8 student would understand." Negotiating with the model is far more efficient than writing from scratch. The best teachers work through 3-4 iterations. One warning: AI models approve of things roughly 50% more than humans do. If you ask "is my lesson plan good?", you'll almost always get praise. If you want criticism, ask for it explicitly: "Tell me the 3 weakest points of this plan. Don't give me praise, just criticism. What student profile would it fail for?"
Pitfalls

Five common mistakes.

1 Writing too generally
Why it's a problemWith prompts like "write a maths question", the model produces randomly across an unlimited space; the result resembles the model's most likely guess, not what you wanted.
SolutionSpecificity isn't about limiting the model, it's about guiding it. Which year group, which topic, how many questions, what format?
2 Giving contradictory instructions
Why it's a problem"Short but detailed", "fun but formal". The model seeks a middle point in the contradiction and manages neither.
SolutionSet a priority. If you want both, give a ratio: 70% engaging, 30% informative.
3 Trusting blindly
Why it's a problemAI can fabricate sources, give wrong dates. It sounds correct because it says it confidently.
SolutionVerify output containing dates, numbers, quotations and names against independent sources. Adding "if you're not sure, say so, don't make it up" helps but isn't sufficient on its own.
4 Not specifying your context
Why it's a problemA prompt without context can produce examples or cultural references that feel foreign or irrelevant to your students.
SolutionState explicitly: "aligned with the national curriculum", "examples from everyday life in the UK". If you give a curriculum code, you anchor the model directly to your local framework.
5 Settling for the first output
Why it's a problemThe most common but most easily fixed mistake. The first output is rarely perfect; the user who thinks "it's fine as is" never sees the prompt's full potential.
SolutionMake iterative refinement a habit. The first output is a beginning, not an end.
Ethics bridge

Watch what's hidden in your prompts.

AI companies need vast amounts of real human data to train models. Every prompt you write could potentially become training data somewhere. Any information from your classroom, your students or parents that enters a prompt is no longer entirely under your control.

Anonymise

Not "Ahmed's latest test" but "a Year 7 student's test". Don't write real names, ID numbers, or class identifiers. If you're uploading a test paper, cut out identifying information.

Check your settings

Look for a setting like "don't use this conversation for training". It exists in most tools but is on by default.

If in doubt, don't share

The simplest and most powerful rule. Don't share a parent's message verbatim; anonymise the subject and consult.

The pedagogical and ethical messages of the workshop require the same attention: write a specific prompt, it improves output; don't give specific personal data, it protects privacy.

Test yourself

Fill in the blanks.

The information the model can hold in mind at the same time is called the
  • token pool
  • context window
  • training data
. If the model forgets early instructions in a long conversation, the reason is, like papers falling off a desk,
  • losing internet connection
  • exceeding the limit
  • the model getting tired
. To receive critical feedback, you should
  • ask for it explicitly
  • hint at it
  • never mention it
.

Why is the prompt "write a short but detailed summary" weak?

Contradictory instructions like "short but detailed" or "fun but formal" should be replaced by setting a priority; if needed, give a ratio: 70% engaging, 30% informative.

Module 2 complete.

You now know how the model sees the world and how to talk to it. In the final module, you'll use this skill to produce a real learning objective from start to finish and assess it two different ways.

Module 3: Materials & assessment →