Module 1 · Lesson 1 of 18Thinking Like a Prompt EngineerFree preview
Thinking Like a Prompt Engineer
Why expert AI users write briefs instead of questions, and the six-part prompt structure that gets strong results on the first or second try.
On this page
You already use AI every week. Some days it hands you something excellent. Other days you get a polite, generic answer that takes longer to fix than it would have taken to write yourself. Most people blame the model, rephrase a little, hit regenerate, and hope. Expert users do something different. They treat every prompt as a brief to a very capable colleague who knows nothing about their situation. When the output disappoints, they diagnose it the way an engineer diagnoses a failing process: by finding the missing input.
This lesson gives you that mindset and a structure you can apply to any task today. Everything else in the course builds on it.
Learning objectives
By the end of this lesson, you will be able to:
- Explain why vague prompts produce generic output, and what the model does with the gaps you leave
- Write a prompt using the six parts of a strong brief: goal, context, task, constraints, format, and examples
- Apply the "new hire test" to spot missing information before you hit send
- Diagnose weak output by its symptoms and fix the prompt instead of rerolling
- Describe the four layers of skill this course develops: prompts, context, workflows, and evaluation
Why vague prompts get generic answers
A large language model (LLM) is the technology behind ChatGPT, Claude, Gemini, and Microsoft Copilot. It writes a response by predicting what text is most likely to come next, given everything it can see. When your prompt leaves a gap, the model doesn't stop and wonder. It fills the gap with the most typical answer for that kind of request.
That's why a short prompt gets a short-prompt answer. Ask for "an email to a client about a project delay" and you get something like the average of every delay email the model has learned from: apologetic, vague, a little stiff. Nothing in it is wrong. It just isn't yours. The model doesn't know who the client is, what caused the delay, what you need the client to do, or how much goodwill you have left.
Tip: The model can only use what's in front of it. Everything it knows about your situation is what you put in the prompt, plus whatever your tool adds behind the scenes. You'll learn to control that second part in Module 3.
Compare these two prompts.
Write an email to a client about the project delay.
Goal: Keep the client's trust after a delay and get their OK on a revised go-live date by Friday.
Context: I'm a project manager at an IT services firm. We're moving a six-location dental group to a new patient scheduling system. Go-live was set for November 3. Their previous software vendor delivered the data export nine days late, so go-live moves to November 17. No data is at risk and testing is on track. My contact is Dana, their operations director. She is direct and busy, and she was frustrated by a delay on an earlier project with us.
Task: Draft the email to Dana explaining the delay and proposing the new date.
Constraints:
- Under 180 words.
- State the cause once, factually. Don't blame the vendor in a way that sounds like an excuse.
- Keep any apology to one sentence.
- Offer one concrete step that lowers her risk, such as a 30-minute walkthrough of the test results.
Format: A subject line, then the email body. After the email, give two alternative subject lines.
The second prompt takes a couple of minutes longer to write. In return, you get a draft you can send with light edits instead of one you have to rewrite. That trade, a little more effort up front for much less cleanup afterward, is the core economics of prompt engineering.
The six parts of a strong brief
The second prompt follows a structure you can reuse for almost any task. Not every prompt needs all six parts, but checking each one takes seconds and catches most gaps.
| Part | The question it answers | Example from the delay email |
|---|---|---|
| Goal | What outcome do you want, and why? | "Get the client's OK on a revised date by Friday." |
| Context | What does the model need to know about the situation, the audience, and the background? | Who Dana is, what happened, what's at stake |
| Task | What exactly should the model produce? | "Draft the email." |
| Constraints | What are the rules: length, tone, must-include, must-avoid? | "Under 180 words. Apology in one sentence." |
| Format | What shape should the output take? | "Subject line, body, then two alternative subject lines." |
| Examples | What does good look like? (Optional but powerful) | A past email you were proud of |
A few notes on using it well:
- The goal is the part people skip most, and it often matters most. "Write a summary" and "Write a summary so my VP can decide by Thursday whether to renew this vendor" produce very different summaries. The goal tells the model what to prioritize and what to leave out.
- Context is where your expertise lives. The model knows a great deal about the world in general and nothing about your client, your company's priorities, or what happened last quarter. You supply that.
- Constraints work best when they say what to do, not only what to avoid. "Use plain language a first-time customer would understand" works better than "Don't use jargon." The next lesson explains why.
- Examples are the fastest way to communicate style. You'll go deep on this in Module 2.
What about opening with a role, like "You are an expert marketing strategist"? A role can help set vocabulary and perspective, and it costs nothing to add. But a role is not a substitute for context. "You are a world-class copywriter" with no information about the product, the audience, or the offer still produces generic copy. It just sounds more confident. Specific context beats an impressive title.
Example: A finance manager asks, "Summarize this vendor contract," and gets a two-page overview she can't use. She rewrites it: "I'm deciding whether to renew this contract before the auto-renewal date. Summarize only the renewal terms, the termination notice period, any price increase clauses, and anything that would penalize us for leaving. Use a table with the clause, the page number, and a one-line plain-English explanation." Same document, completely different usefulness.
The new hire test
Here's a quick check you can run on any prompt before you send it:
If you handed this prompt to a smart new hire on their first day, with no access to your head, your files, or your history, could they do excellent work?
If the answer is no, ask yourself what they would come back and ask you. "Who is this for?" "How long should it be?" "What did we promise them last time?" Each of those questions points to missing context. Answer them in the prompt.
When you're not sure what's missing, let the model run the new hire test for you:
I need your help with the task below. Before you start, ask me up to five questions about anything that would meaningfully change your answer: the audience, the purpose, constraints, background, or what good looks like. Ask them all at once, numbered. Don't start the task until I've answered.
Task: [describe the task]
This works especially well for unfamiliar or high-stakes work, where you don't yet know what you don't know. Over time, it also teaches you which context matters for which kind of task, so you start including it without being asked.
Iterate deliberately instead of rerolling
When output disappoints, the casual move is to regenerate and hope for a better roll of the dice. Sometimes that works, because model output varies from run to run. But it doesn't fix the cause, so the problem comes back next time.
The expert move is to diagnose. Look at what's wrong with the output and trace it back to what was missing or unclear in the input.
| Symptom | Likely cause | Fix |
|---|---|---|
| Generic; could be about anyone | Missing context | Add specifics about the situation, the audience, and the stakes |
| Wrong tone | Audience or voice not described | Describe the reader and your relationship with them; add a sample of the voice you want |
| Too long or too short | No length constraint | Give a word count, a number of bullets, or "fits on one screen" |
| Misses the point | Goal unclear | State the purpose and the decision or action the output should support |
| Ignores an instruction | Instruction buried or contradicted | Move it up, make it explicit, and remove conflicting instructions |
| States "facts" that are wrong | Asked for information it doesn't have | Provide the source material and allow it to say "not stated" |
| Format you can't use | No format specified | Ask for a table, bullets, named sections, or the exact structure you need |
Then make one or two targeted changes and run it again. If you change five things at once, you won't know which one helped.
Most important: fix the prompt, not just the output. When you discover that adding "my reader is a CFO who skims" transforms your summaries, put that line into the prompt you saved, not just into this one chat. Every fix you capture makes the next run better. That habit is the seed of the reusable template library you'll build in Module 4.
Watch out: Correcting the model inside a long conversation ("No, shorter. No, more formal. No, the other client.") works for the task in front of you, but those lessons disappear when the chat ends. When a correction matters, copy it back into your saved prompt.
What "prompt engineering" means in this course
Prompt engineering sometimes gets described as a hunt for magic words. It isn't. Today's models understand plain language well, so the leverage has moved from clever wording to four things you control:
| Layer | The question it answers | Where you'll learn it |
|---|---|---|
| Prompts | Are my instructions clear, complete, and well structured? | Modules 1 and 2 |
| Context | Does the model have the right information, and nothing that distracts it? | Module 3 |
| Workflows | Is the work broken into steps, packaged for reuse, and repeatable by others? | Modules 4 and 5 |
| Evaluation | Do I know it works, and can I show the value? | Module 6 |
Think of the difference between someone who can cook a great meal once and a restaurant kitchen that serves the same quality every night. This course builds the kitchen: prompts you can trust, context you control, workflows you can hand to a colleague, and evidence that they work.
Try it
Spend 10 to 15 minutes on this. It's the starting point for your personal prompt library.
- Pick a task you've done with AI in the past two weeks where the result was "fine but generic." A summary, an email, a plan, a product description, anything.
- Rewrite your original prompt using the six parts: goal, context, task, constraints, format, and an example if you have one handy.
- Run the new hire test on your rewrite. Write down any question a new hire would ask, and answer it in the prompt.
- Open two fresh chats. Run your original prompt in one and your new prompt in the other.
- Compare the outputs. Note what changed, and estimate how many minutes of editing each would need before you could use it.
- Create a document called Prompt Library and save your improved prompt in it, with a one-line note on what it's for. You'll add to this file throughout the course.
If the new version isn't clearly better, use the symptom table to find the likely cause, make one change, and run it again.
Key takeaways
- The model fills every gap you leave with the most typical answer, so vague prompts produce generic output.
- A strong prompt is a brief: goal, context, task, constraints, format, and examples. The goal and the context usually do the most work.
- A role can help, but specific context beats an impressive title.
- Use the new hire test, or ask the model to question you first, to find missing information before you send.
- Diagnose weak output by its symptoms, change one or two things at a time, and save every fix back into your prompt.
- Prompt engineering is about prompts, context, workflows, and evaluation, not magic words.
Enjoyed this preview?
Unlock all 18 lessons, 3 templates, and your certificate for $249.