Lesson 2 of 5 · The Art of Prompt Engineering
Lesson 2
The five principles of a good prompt
Specificity. Context. Role. Format. Constraints. Five principles that, applied consistently, turn a mediocre prompt into a useful one. We will improve a single recurring task — a weekly newsletter — through this lesson, in five small steps.
By the end of this lesson, you will:
- Know the five principles that improve almost any prompt: Specificity, Context, Role, Format, Constraints.
- Have seen each principle applied as a single small improvement to a real task, with before-and-after results.
- Have rewritten a prompt of your own using all five principles and seen the output change.
The task we will improve
Throughout this lesson we work on one task. The goal is to write a weekly internal newsletter for our team summarising the past week. It is the kind of task many people do, in many forms, in many organisations — and the kind that AI is well suited to help with.
We start with the most natural prompt a beginner would write.
v0 — The starting prompt
Write a weekly newsletter for my team.
You can try this. The result will be a generic template that does not know what team, what week, what news, or what tone. Useful as a skeleton, but it is not yet useful as a draft. Let us improve it, one principle at a time.
Principle 1 — Specificity
The single most powerful improvement to any prompt is being specific about what you want. Vague prompts produce vague outputs. The model has been asked to fill in vast gaps and will do so based on what it imagines is most likely. Often it is not what you imagined.
Apply it to our newsletter.
v1 — With specificity
Write a weekly newsletter for my team of 12 software engineers working on a customer-facing payments product. Cover the work done this week, anything coming up next week, and any notable news from outside our team that they should be aware of.
Already much better. The model now knows who the team is, what they work on, and what topics should be in the newsletter. But the response is still going to be largely made up — the model does not know what was actually done this week. We fix that next.
Principle 2 — Context
The model is not in the room with you. It does not know what happened in your meetings, what the team built last week, what is on the calendar. If you want the newsletter to reflect reality, you have to provide the reality.
v2 — With context added
Write a weekly newsletter for my team of 12 software engineers working on a customer-facing payments product. Here is what happened this week: - Shipped the new dispute-handling flow (3-week project, on time) - Started the migration of the legacy fraud-check service - Two new hires (Sarah and Mateo) joined on Monday - One incident on Wednesday — payment-confirmation emails delayed by ~20 minutes for about 4,000 customers; resolved in 90 minutes; retro tomorrow Coming up next week: - Code-freeze for the quarterly release on Thursday - Sarah and Mateo start their first features - All-hands on Friday Cover the work done, what is coming up, and a brief note on the incident. Highlight the new joiners.
Now the prompt has the substance the model needs. The response will be a real first draft of a real newsletter. But it will still be in whatever voice the model defaults to — possibly too breezy, possibly too corporate. We can shape that.
Principle 3 — Role
Telling the model who it is helps it adopt the voice and stance you want. The "role" instruction is not magic; it is a way of giving the model a strong hint about the kind of writing you expect.
v3 — With role
You are the engineering manager of a 12-person team. You write a weekly newsletter that is direct, warm, and respectful of your team's time. You do not use marketing language. You do not over-celebrate routine work. You acknowledge problems honestly. Your team values clarity and brevity. [same content and instructions as v2 follow]
This single addition often changes the response noticeably. The model is now more likely to write something that sounds like a competent manager rather than a generic AI assistant. The tone, the structure, the level of corporate-ness — all are nudged in the direction you want.
Principle 4 — Format
Telling the model how the response should be structured saves you the work of restructuring it yourself. Most prompts benefit from explicit format guidance.
v4 — With format
[same role and content as v3] Structure the newsletter as follows: - Subject line (short, no clickbait) - One-paragraph opening that frames the week - "What we shipped" — bullet list, one line each - "Coming up" — bullet list, one line each - "A note on Wednesday's incident" — two sentences, honest - "Welcome" — one paragraph introducing Sarah and Mateo - Sign-off Total length: 250-350 words.
The model now produces output you can almost paste directly. The structure is yours, the substance is yours, the model has done the writing.
Principle 5 — Constraints
The final principle is telling the model what not to do. Constraints are often more useful than positive instructions — they head off the model's default tendencies that you do not want.
v5 — With constraints
[same role, content, and format as v4] Constraints: - Do not use the word "exciting". - Do not use exclamation marks anywhere except in the welcome paragraph. - Do not speculate about the cause of the incident — the retro is tomorrow and we have not yet established it. Stick to what is known (delay, scope, resolution time). - Do not include emoji. - Do not use AI-cliché phrases like "I'm thrilled to share" or "without further ado". Direct, professional voice only.
This final version is, in practice, the difference between a prompt that produces a draft you have to heavily rewrite and a prompt that produces a draft you only need to edit lightly. The constraints address the specific tendencies of AI-generated writing — the over-celebration, the cliché phrasings, the inappropriate certainty — that would otherwise require manual cleanup.
The five principles, on one page
The five principles, summarised
1. SPECIFICITY — Be precise about what you want.
Bad: "Summarise this."
Good: "Summarise this 12-page report in 200 words,
focused on the three recommendations."
2. CONTEXT — Give the model the facts it needs.
The model is not in the room with you. Paste in
the relevant data, prior emails, source documents.
3. ROLE — Tell the model who it is.
Not for magic. To nudge it toward a particular
voice, register, and stance.
4. FORMAT — Tell the model how to structure the output.
Headings. Bullet lists. Tables. Word counts.
Saves you restructuring.
5. CONSTRAINTS — Tell the model what NOT to do.
Words to avoid. Tones to avoid. Assumptions
to avoid. Often the most-useful principle.
Aside · You do not always need all five
For a quick question, "specificity" is often enough. For drafting that has to land in a particular voice, you need all five. For batch tasks that you will repeat (drafting many similar emails, processing many documents the same way), investing twenty minutes in a five-principle prompt pays back across every future use of it. Calibrate the effort to the importance of the output.
Exercise — Apply the five principles to your own task (25 minutes)
- Pick a real task you do regularly that AI could help with. Examples: drafting a meeting agenda, replying to a difficult email, summarising a long document, writing a project status update, planning a lesson if you teach.
- Write the natural beginner's prompt for it — one sentence. Run it. Note the result.
- Now add the five principles, one at a time, running the new prompt after each step. Specificity → Context → Role → Format → Constraints.
- After all five, save the final prompt somewhere you can reuse it. This is your first entry in a personal prompt library (we come back to this in Lesson 4).
- Reflect: which of the five made the biggest difference for your task? Different tasks benefit most from different principles. Knowing which you tend to under-use is a working insight.
Self-check
- Name the five principles in order.
- Why is "role" not magic — what is it actually doing?
- Why are constraints often more useful than positive instructions?
- When is the natural beginner's prompt enough, and when do you need all five principles?
Looking ahead
Lesson 3 introduces the specific patterns — few-shot examples, chain-of-thought, structured output, decomposition — that go beyond the five principles. These are the techniques that have been studied and documented in the research literature; we will show how to apply them in everyday work.