Lesson 1.1: What a prompt is and how a model uses it

Learn what a prompt really is, what a language model can and cannot know, and why a sharper prompt saves you time on every task you give it.

By the end of this lesson you will be able to explain, in plain words, what a prompt is and what a language model does with it, list the things that count as part of a prompt, and say why the same request can give a great answer one day and a useless one the next. You will use a simple test, the new-colleague test, to find the gaps in any prompt before you send it, and you will see with numbers why a vague prompt costs more time than a careful one. You need no technical background and no particular AI tool. Everything here applies to any assistant you use, and the exercises can be done by pasting text into whichever one you prefer.

Many people meet AI tools the way they meet a search engine: type a few words, read what comes back. When the result is poor, they conclude the tool is poor. Often the tool is fine and the prompt was thin. Prompt engineering is the skill of closing that gap, and it begins with understanding what the tool actually does with your words.

What a prompt is

A prompt is everything the model reads before it writes its reply. It is more than the line you type. Depending on the tool, it can include several layers, and all of them shape the answer. Part of the prompt · What it is · Who usually writes itSystem instructions · Standing rules about role, tone and limits · The app builder, or you in a settings page Your message · The request you type now · You Conversation history · Earlier turns in the same chat · You and the model, automatically included Attached material · Documents, spreadsheets, pasted text or images · You Tool results · Search results or data fetched during the answer · The tool, behind the scenes The model has no memory of you outside this text. If something is not in the prompt, in the history or in what the model learned during training, it does not know it. This is the single most useful fact in the whole course, and most prompting advice follows from it.

What the model does with it

A language model is trained on a very large amount of text to do one job: given some text, predict what is likely to come next. When you send a prompt, it produces a reply one small piece at a time, each piece chosen because it is a plausible continuation of everything so far. Think of an extremely well-read colleague who has absorbed a huge library, writes fluently, and has no idea who you are, what your company does or what you meant, beyond what you tell them.

Three consequences follow, and each one explains behaviour that surprises beginners.

It writes what is plausible, not what is checked. A fluent, confident reply is not evidence of a correct one. The model can produce a convincing sentence about a policy that does not exist, because a convincing sentence was a likely continuation. Lesson 2.4 covers how to catch this.

It fills gaps with guesses. If you leave something unspecified, the model picks the most likely option. If you ask for "a summary of the report" it picks a length, an audience and a tone for you. Sometimes the guess is right. Often it is average, which is rarely what you need.

The reply can differ each time. Models choose among likely continuations with some randomness, so the same prompt can give different answers. That variation is a feature when you want ideas and a problem when you want consistency. Lesson 6.1 returns to making results consistent.

Predict first: you type "Write an email to a client about the delay." What do you think the model has to guess, because you did not say? List at least five things before reading on.

A fair list includes who the client is, how long the delay is and why it happened, how formal to be, how long the email should be, whether to apologise, whether to offer something, what you want the client to do next, and your name and role. That is eight or nine unknowns in one short sentence. The model will guess all of them. Each guess may be reasonable and the result may still be wrong for your situation.

The new-colleague test

Here is a test you can apply to any prompt in a few seconds. Imagine handing exactly your message, and nothing else, to a smart, willing person who has just joined your team. They know nothing about the project, the client or your preferences. Could they do the task well from your message alone? If they would need to ask five questions first, the model needs those answers too, and it cannot ask unless you invite it to.

The test is useful because it turns a vague feeling ("the AI is not getting it") into a concrete list: the missing facts. Take the email prompt again.

> Write an email to a client about the delay.

A new colleague would ask: which client, delayed by how much, what happened, what do you want to say about next steps, how formal? Now a version that passes the test.

text Write a short, polite email from me (Meena, project manager at a payments company) to Mr Rao, the finance head at a client bank. Our integration go-live, planned for 14 November, will move to 28 November because the client's security review took longer than planned. Say sorry for the change without blaming anyone, confirm that all other milestones are unchanged, and ask him to confirm the new date by Friday. Keep it under 120 words, in a warm but professional tone.

The second prompt is about 90 words, and it takes under a minute longer to write. It removes nearly every guess. Notice what it did not do: it did not use clever tricks. It simply supplied what a colleague would need.

Why vague prompts cost more, not less

A short prompt feels efficient. The trap is that its cost shows up later, in rewrites. Each unspecified detail is a guess, and each guess has some chance of being wrong. If you leave several details open, the chance that all the guesses fit your needs drops fast.

Suppose that for each missing detail the model has a 60 per cent chance of guessing what you wanted. With one missing detail, the first answer is right 60 per cent of the time. With four missing details, all four must be right, so the chance is 0.6 × 0.6 × 0.6 × 0.6 = 0.13, or 13 per cent. In other words you should expect to ask roughly eight times before everything lines up, or, more realistically, you will give up and edit by hand. The numbers are illustrative, and real details are not independent, but the pattern is real: vagueness compounds.

Now count the time. Suppose the short prompt takes one minute to write, and you then need about four rounds of reading and correcting at three minutes each, so 1 + 4 × 3 = 13 minutes. The careful prompt takes four minutes to write, and you need one round of reading and light editing, so 4 + 1 × 3 = 7 minutes. The careful prompt saves about six minutes on a single email. Over a week of such tasks, the habit is worth hours. Try your own numbers.

There are real exceptions. For a quick question where any reasonable answer will do, a short prompt is fine. The skill is knowing which kind of task you are doing: when the details matter, say them.

What the model cannot know

It helps to keep a short mental list of things that are almost never in the model's head unless you put them in the prompt.

Your situation: your company, your clients, your team, your constraints. Your preferences: how formal you like to sound, which words you avoid, what you consider "short". Private or recent facts: last week's numbers, an internal policy, a decision made in yesterday's meeting. Many tools can search the web or read files you attach, but only if you let them and only for what they find. The purpose behind the request: a summary for a CEO and a summary for a compliance auditor are different documents. What good looks like: the model cannot see your standard unless you describe it or show an example.

The remedy is the same each time: put the missing facts in the prompt, or attach the material. The next lessons give you a structure for doing this quickly.

Reading a reply critically: the five-point first read

The new-colleague test improves what you send. The other half of the skill is how you read what comes back. Treat every reply as a draft from a capable but unchecked colleague, and make the same first pass each time.

1. Did it answer the question you asked, or a nearby, easier one? Models sometimes answer what is most common instead of what you wrote. 2. Are the facts supplied by you, or invented? Mark every name, number, date and quotation. Anything not in your prompt or your attachments needs checking. 3. Is anything missing that you expected? Silence is easy to miss. If you asked for three risks and got two, count. 4. Does the tone and length fit the reader? A reply can be correct and still unusable because it is too long or too chatty. 5. What assumptions did it make? Ask the model directly: "List the assumptions you made in this answer." The list often shows the exact gap in your prompt.

The fifth step is worth practising, because it turns every weak reply into a diagnosis. Suppose you asked for an email about a delay and the reply apologised heavily and offered a discount you never mentioned. The assumptions list will say "assumed the client is upset and that compensation is appropriate". Now you know which sentence to add: "Do not offer compensation." Over time you will see the same few gaps again and again, and your first prompts will start to close them before you send.

A short example of the layers at work

It helps to see the layers from the table together. Imagine a bank gives its relationship managers an internal assistant. The system instructions, written once by the team that built it, say: "You help relationship managers prepare for client meetings. Use a professional tone. Never state interest rates unless they appear in the attached rate sheet." A manager then attaches the rate sheet and the client's profile and types a message: "Prepare me for tomorrow's call with Mr Rao about his working-capital limit."

By the time the model replies, its prompt contains the standing rules, the rate sheet, the client profile, the message, and any earlier turns of the chat. It never saw the manager's calendar or the previous week's phone call, so it cannot refer to them unless the manager adds them. If the manager had typed only "Prepare me for tomorrow's call", the model would have had no idea which client, and a well-built assistant would ask. A poorly built one would invent a client, which is the plausible-continuation behaviour at work. Two lessons follow. When a tool behaves oddly, ask what is in its prompt that you cannot see. And when you are the one writing standing instructions, you are writing the system layer for everything that follows.

Different tools, same idea

Assistants differ. Some remember earlier chats, some can browse the web, some can run code, and some let you set standing instructions or upload files. The details change every few months, so this course does not tie itself to one product. What does not change is the basic situation: a model reads text, including text you attached, and writes the most fitting continuation. Whenever a tip in this course depends on a feature, we will say so. And because tools update, always test a technique on the tool you actually use before you rely on it.

Worked examples: the new-colleague test in practice

Here are three everyday requests, the questions a new colleague would ask, and a rewrite.

Example 1: summarising. The first draft is "Summarise this report." A colleague would ask who it is for, how long it should be, and what to focus on. A better prompt is: "Summarise the attached credit-risk report for our head of operations in five bullet points. Focus on what changed since last quarter and any action we need to take. Do not include methodology." The gaps closed are audience, length, focus and exclusions.

Example 2: analysis. The first draft is "Analyse our sales." A colleague would ask which sales, over what period, compared with what, and what decision depends on it. A better prompt is: "I have monthly sales by region for the last 12 months (pasted below). Identify the two regions with the weakest growth and suggest three possible reasons I should check, based only on the data. Do not invent figures." Notice the last sentence: it sets a boundary against making things up.

Example 3: a policy question. The first draft is "Can employees carry forward leave?" A colleague would ask which company, which country and which policy document. A better prompt is: "Using only the leave policy pasted below, answer whether an employee can carry forward unused leave, and quote the sentence that supports your answer. If the policy does not say, reply 'The policy does not specify.'" This ties the answer to a source and gives the model a safe way to say it does not know.

In all three the improvement was not extra length for its own sake. Each added sentence closed a gap that would otherwise have been guessed.

Common mistakes

Assuming the model knows your context. It knows only what is in the prompt and in its training. Say who, what, why and for whom. Treating fluent as correct. Polished prose is not verification. Check facts that matter. Writing the prompt like a search query. A few keywords is a weak prompt. Write as you would brief a colleague. Over-trusting a first answer. The first reply is a draft. Read it critically and ask for changes. Never saying what good looks like. Give a length, a format, an audience or an example. Blaming the tool before checking the prompt. Run the new-colleague test first. Pasting sensitive data without thinking. Do not paste customer data, passwords or confidential documents into a tool your organisation has not approved. Lesson 6.3 covers this.

Key takeaways

A prompt is everything the model reads before replying: system instructions, your message, earlier turns, attachments and tool results. The model predicts a plausible continuation. It does not check facts, it fills any gap with a guess, and its answers can vary from one run to the next. If something is not in the prompt or in the model's training, it does not know it, including your situation, your preferences and your private or recent facts. The new-colleague test finds the gaps: could a smart newcomer do this task from your message alone? Each open detail is a guess, and guesses compound. With four open details at 60 per cent each, the first answer fits about 13 per cent of the time. A careful prompt costs minutes to write and usually saves more minutes in rewrites. For trivial questions a short prompt is fine.

Check yourself

1. What are five things that can be part of a prompt, besides the line you type? System instructions, earlier conversation, attached files or pasted text, images, and tool results such as search findings. 2. Why can a model give a confident answer that is wrong? It generates a plausible continuation. It does not check facts, so fluency and correctness are separate. 3. If each of three open details is guessed right 70 per cent of the time, what is the chance all three fit? 0.7 × 0.7 × 0.7 = 0.343, about 34 per cent. 4. You send "Summarise the report." Name three details the model must guess. The audience, the length and format, and what to focus on. Also the tone and what to leave out. 5. A short prompt takes 1 minute plus 4 rounds of 3 minutes. A careful one takes 4 minutes plus 1 round of 3 minutes. How much time does the careful prompt save? 13 − 7 = 6 minutes per task.

Try this

This exercise needs any AI assistant and about thirty minutes. You will compare a thin prompt with a careful one on a real task, and score both.

1. Pick a task you did this week or will do soon: an email, a summary, a short analysis, a plan. Write it as you would naturally type it, in one line. Send it, and save the reply. Write "Version A" above it. 2. Apply the new-colleague test to your own line. List every question a newcomer would ask. Aim for at least five. 3. Write "Version B": a prompt that answers those questions, in under 150 words. Send it in a fresh chat, so the earlier reply does not influence it, and save the reply. 4. Score both replies against this rubric, giving each question one point if the answer is yes. Question · A · BIs the audience right? · · Is the length right? · · Is the tone right? · · Are the facts all ones you supplied or can verify? · · Could you use it with only light edits? · · 5. Count how many rounds of correction each version would need before you could use it, and use the time simulator above with your own numbers. 6. Write three sentences: what the thin prompt left to guesswork, what the careful prompt fixed, and the one detail that mattered most.

A good result is a pair of replies with clearly different scores and a short note naming the missing details. Keep the careful prompt: in Lesson 1.3 you will turn it into a reusable template. In the next lesson we look at how the model reads your words, in units called tokens, and why the amount of text you send matters.

Tags: Prompt Engineering Prompts Language models AI basics Beginner