Why prompts change everything
Remember attention: every token in your prompt gets consulted while the model writes every token of its answer. A prompt with a clear role, audience, and format literally reshapes the probabilities of every output token. "Write about photosynthesis" and "You are a CSEC biology tutor; explain photosynthesis to a grade 10 student in 5 bullet points with one Jamaican example" travel through the same network, but they light up very different paths.
From Lesson 1: the model picks each next token by rolling weighted dice, and the weights come from everything already in the context. Your prompt IS that "everything already there", so it re-weights the dice for every single token that follows. Feel it with two tiny prompts. After "My favourite city is", thousands of cities share the probability: Paris, Tokyo, Miami, Kingston, all plausible. But after "I grew up near Half Way Tree, so my favourite city is", the token "Kingston" rockets to near-certain, and every sentence after that will lean Jamaican too. You did not command the model; you shifted the probabilities so hard that the answer you wanted became the most likely one. That is all prompt engineering is: moving probability, on purpose.
The set of all possible continuations the model is choosing between, each with its own likelihood. A vague prompt leaves this space huge (millions of generic answers are equally likely). Every specific detail you add shrinks it toward the answers you actually want.
The five-part framework
Who should the AI be? "You are an experienced CSEC maths tutor" summons different knowledge than "you are a stand-up comedian".
What does it need to know? Audience, situation, constraints: "for a 15-year-old in Trinidad preparing for exams in 3 weeks".
One clear verb-first instruction: "Create a 7-day revision plan". Not "maybe help with some studying stuff?".
What should the output LOOK like? "A table with columns Day, Topic, Practice", "under 200 words", "JSON".
Show one sample of what good looks like ("few-shot" prompting). The single biggest quality upgrade for tricky tasks.
The first output is a draft. "Make it shorter", "row 3 is wrong, fix it", "more local examples". Treat it as a conversation, not a slot machine.
Watch one prompt level up: bad, good, great
Here is the framework in action on one real task: getting help to prepare a presentation on Jamaican Carnival for your school's Culture Day. Same model, three prompts, three very different outputs.
- Bad: "tell me about carnival". Zero constraints, so the probability space is enormous. The model averages over everything it has seen labelled "carnival": Rio, Trinidad, cruise ships, Notting Hill. You get a generic Wikipedia-flavoured blob that mentions Jamaica maybe once. Not because the model is lazy, but because you left it guessing which of a million possible answers you meant.
- Good: "Explain the history of Carnival in Jamaica for a 10-minute school presentation to teenagers." Adding the task (explain history), the audience (teenagers), and the situation (10-minute presentation) collapses most of the wrong answers. Rio disappears. The length becomes sensible. Each detail you added killed off a whole family of unwanted continuations.
- Great: add role, format, and an example. "You are a Caribbean cultural historian who makes teens care about heritage. I am presenting Jamaica Carnival history at Culture Day. Give me a 10-minute outline as a table: Section, Key point, Hook for the audience. Here is the style I want for one row: 'Origins | Carnival reached Jamaica via Trinidad's traditions plus our own Jonkonnu roots | Play a 5-second soca vs Jonkonnu drum clip.'" The role pulls the answer toward historian-grade knowledge instead of tourist-brochure fluff. The format makes the output instantly usable. The example row is the strongest signal of all: the model continues patterns, so showing one perfect row makes every other row imitate it.
Notice what each upgrade did. Role changed WHICH knowledge got surfaced. Context and task changed WHAT the answer is about. Format changed the SHAPE. The example changed the STYLE of every line. None of it was magic; all of it was probability-narrowing.
🎛️ The prompt builder
Build a prompt for a real scenario: getting help with a school project. Add parts one at a time and watch the prompt (and its grade) grow. Read the final version out loud; you will hear the difference.
Two students ask for essay help. A: "help with my essay". B: "You are an English teacher. My essay argues for renewable energy in Barbados, audience is my CSEC examiner, 500 words. Point out my 3 weakest arguments and suggest evidence for each." Why does B get dramatically better output?
You will see posts claiming secret phrases like "take a deep breath" or "I will tip you $200" unlock hidden power. It is a tempting belief, because it feels like finding a cheat code, and sometimes those phrases do nudge results a little (they resemble contexts in the training data where answers were more careful). But that is the whole trick: they are just more context tokens shifting probabilities, exactly like every other word you type. There is no locked "smart mode" waiting for a password. A prompt with a clear role, real context, and one good example will beat any magic phrase every time, because clear context narrows the probability space far more than a superstition does. Skip the incantations; write the constraints.
When the answer is wrong or generic: the debugging routine
In Week 5 you debugged models by checking the data and the loss. Prompts get debugged the same way: methodically, not by rage-typing "no, do it BETTER". Run these steps in order:
- 1. Diagnose the miss. Is the answer wrong (facts are off), generic (right topic, no depth), or off-target (answered a different question)? Each failure has a different fix, so name it first.
- 2. Generic answer? Add context. Generic output means your prompt matched a million contexts. Add who you are, where you are, and what you already know: "I am in 4th form in Jamaica and already understand X" removes the beginner boilerplate.
- 3. Wrong answer? Challenge and verify. Ask "check that step again" or "what is your evidence for row 3?", and confirm important facts outside the chat. Never fix a factual error by just regenerating and hoping.
- 4. Wrong shape? Show, don't describe. If the format is off, paste one example of exactly what you want. One sample beats three paragraphs of format instructions.
- 5. Still stuck? Restart the conversation. Everything in the chat is context, including the bad answers, and they keep steering new output toward the same failure. A fresh chat with your improved prompt clears the polluted context.
Prompting an AI to teach, plan, critique, and quiz you is a superpower. Prompting it to write your SBA so you can paste it in is plagiarism, and examiners (and CXC's AI rules) are increasingly good at catching it. The framework you just learned is for the first kind. Use it to get smarter, not to outsource being you.
1) The prompt is the steering wheel: every token shapes every output token via attention. 2) Build prompts from Role + Context + Task + Format + Examples. 3) Iterate: the first answer is a draft, not a verdict. 4) Use it to learn, not to cheat.
Next: the hard questions. Deepfakes, bias, and power. ⚖️