"The model is only as clear as your request. Prompting well is the fastest, cheapest way to get dramatically better AI output."
Level: Beginner · Time: ~13 min · Prerequisites: Module 12
Learning Objectives
By the end of this module, you will be able to:
- Write clear, effective prompts
- Use the four parts of a strong prompt
- Apply techniques like few-shot and step-by-step
- Improve prompts through iteration
1. A Prompt Is a Brief
A prompt is the instruction you give an LLM. Prompting well is less about secret tricks and more about clear communication — like briefing a capable new assistant.
Strong prompts usually have four parts:
- Instruction — what to do ("Summarize…", "Write…")
- Context — who it's for, background
- Examples — show the style you want
- Constraints — length, format, tone
Explain like I'm new: "Write about dogs" is a vague brief and gets a vague result. "Write a friendly 100-word intro for first-time dog owners, warm tone, one tip, end with a question" is a clear brief — and the output transforms.
Think about how you'd hand a task to a smart new colleague. You wouldn't just say "handle the report" — you'd say what report, for whom, by when, and in what format. The LLM has no way to read your mind or ask a clarifying question mid-task, so every detail you don't give it becomes a detail it has to guess. Prompting is simply the discipline of guessing less on the model's behalf.
2. Bad Prompt vs. Better Prompt
| ❌ Vague | ✅ Improved |
|---|---|
| "Fix my email." | "Make this email more polite and under 5 sentences, friendly tone: [paste]" |
| "Explain APIs." | "Explain APIs to a beginner using a restaurant analogy, under 120 words." |
The improved versions add audience, format, length, and tone — same model, far better output.
Key idea: Adding specifics (audience, format, constraints) is the single biggest lever in prompting. You're removing the model's guesswork and pointing it at exactly the target you want.
3. Reliable Techniques
- Few-shot: show 1–3 examples of the input→output you want. Powerful for steering style/format. (Giving zero examples and just asking is called zero-shot.)
- Step-by-step: "think through this step by step" often improves reasoning and accuracy — nudging the model to show its work instead of blurting a final answer. This is often called chain-of-thought prompting.
- Role: "You are a patient tutor" sets tone and depth.
- Format: ask for a table, bullet list, or JSON — especially useful when another program will read the output.
Real-world use case: In an app, developers write a system prompt once ("You are a support agent for a bakery; only answer from these facts…") that shapes every user interaction. Prompt design is a core skill in building LLM products — and LangChain (next modules) helps manage prompts at scale.
4. Iterate
Prompting is a loop: write, see where the output falls short, fix that one thing, run again. Pros rarely nail it first try — they refine. Treat it like debugging code: if the output is too long, add a length limit and rerun; if the tone is off, name the tone and rerun. Because you change one thing at a time, you always know which tweak helped — and you slowly build a personal library of phrasings that reliably work.
Common mistake: Rewriting the whole prompt from scratch when one detail is off. Change a single variable — add an example, tighten a constraint — and see what improves. You'll learn faster and keep what worked.
Hands-On: Try This
Try this: Take a vague request you'd normally type and rewrite it adding three things: who it's for, how long, and what format. Run both versions and compare. The upgrade is almost always obvious.
✅ Checkpoint
- Name the four parts of a strong prompt.
- What is few-shot prompting?
- What's the right way to improve a prompt that's slightly off?
Answers: 1) Instruction, context, examples, constraints. 2) Showing the model a few examples of the desired input→output. 3) Change one variable at a time and iterate.
Key Takeaway: A prompt is a clear brief — instruction, context, examples, constraints. Adding audience, format, length, and tone dramatically improves output. Use few-shot examples, step-by-step reasoning, and roles; in apps, a system prompt shapes every interaction. Improve prompts by iterating one change at a time.
Further Learning
Part of "Zero to AI Engineer." Adapted from Microsoft's open curricula (MIT License).