Chapter 3 — Prompt Engineering & Context Crafting¶

Part I — AI foundations for developers

In 30 seconds¶

The core idea: good prompts give Copilot clear intent and the right context; how you structure a request materially changes the output.
Why it matters: prompt and context crafting is the highest-leverage skill for day-to-day Copilot use.
The exam angle: GH-300 tests prompt structure and context, how context is determined, zero-shot and few-shot prompting, best practices, and prompt process flow / chat-history usage.
Remember: be specific, provide context, show examples (few-shot), iterate.

Exam map¶

Exam map — GH-300 · Apply prompt engineering and context crafting

1. Key concepts¶

Chapter 2 showed that Copilot builds a prompt from context before the model runs. Prompt engineering is the skill of steering that context and your instructions so the model's probabilities bend toward the code you actually want. GH-300 tests it directly: prompt structure and context, how context is determined, zero-shot versus few-shot prompting, best practices, and how the chat conversation carries history.

📖 Definition — Prompt engineering: the practice of designing and refining the instructions and context you give an AI — what you ask, how you phrase it, and what you reference — to get more accurate, relevant, and useful output.

📖 Definition — Context: everything the model "sees" for your request — your explicit words plus the code Copilot gathers (current file, selection, neighboring tabs) and anything you deliberately reference. Weak output is very often a context problem, not a model problem.

Anatomy of a strong prompt¶

A good developer prompt tends to carry four things. You will not always spell out all four, but the higher-stakes the task, the more each one matters:

Intent (goal) — what you want produced ("write a function that…", "refactor this to…").
Context — the relevant code, constraints, and background ("this runs in a browser, no Node APIs").
Specifics — inputs, outputs, edge cases, libraries, and style ("return a Result, handle empty input").
Examples — a sample input/output or a pattern to imitate.

📌 Key concept: be specific and give the model something to imitate. "Make this better" gives the model nothing to optimize for. "Extract the validation into a pure function, add JSDoc, and keep the public signature unchanged" gives it a target.

Zero-shot vs few-shot¶

📖 Definition — Zero-shot prompting: asking for the result with no example — you rely on the model's general training ("Write a regex for a UK postcode").

📖 Definition — Few-shot prompting: including one or more examples of the desired input/output or style so the model imitates the pattern ("Given these three sample rows, generate ten more in the same shape"). Few-shot is the go-to when format or style must be exact.

2. How it works¶

How Copilot determines context¶

Copilot assembles context from several sources, and you can influence each:

Implicit context — the current file around your cursor, your active selection, and related open tabs (Chapter 2). Managing which files are open is itself a form of prompting.
Explicit references — in Copilot Chat you can point at specific context: chat variables such as #file, #selection, or the whole workspace; participants such as @workspace; and / slash commands such as /explain, /fix, /tests. (Exact syntax varies by IDE — verify in your editor's docs.)
Conversation history — in a chat thread, earlier turns remain in context, so follow-ups like "now add error handling" build on what came before.

🔍 How it works: prompting is a loop, not a single shot. Because generation is probabilistic (Chapter 1), the reliable path is: draft a focused prompt → read the result → refine (add a constraint, an example, or a reference) → repeat. The conversation carries context, so refining usually beats starting over.

flowchart LR
    A["Draft prompt<br/>(intent + context + specifics)"] --> B["Read the response"]
    B --> C{"Good enough?"}
    C -->|No| D["Refine: add example,<br/>constraint, or #reference"]
    D --> B
    C -->|Yes| E["Accept / reuse as a prompt file"]

Best practices that reliably help¶

One job per prompt — ask for a single, well-scoped thing; chain follow-ups rather than cramming.
Give clear intent and specifics — name inputs, outputs, edge cases, and the library or style to use.
Provide examples (few-shot) when the format matters — sample data, a target signature, a style to copy.
Keep relevant files open — you are supplying context (Chapter 2).
Iterate — treat the first answer as a draft; "make it shorter," "handle nulls," "add a test."
Follow good coding practice in the prompt — clear names and comments in your code steer better output.

🖥️ Hands-on: in Copilot Chat, scope the request explicitly and ask for one thing:
/tests #selection
Write unit tests for the selected function. Cover empty input, a single element,
and a duplicate-key case. Use the project's existing test framework.

3. In the real world¶

Scenario — from vague to precise. A developer types // sort the list and gets a generic sort that ignores the domain. Reframed with intent, specifics, and an example, the prompt becomes:

"Sort orders by priority descending, then by createdAt ascending. priority is one of low|medium|high. Return a new array; don't mutate the input. Example: a high order from yesterday should come before a high order from today."

Copilot now produces a comparator with the right tie-breaker and an explicit priority ordering — because the prompt supplied the intent, the constraints, and a concrete example to imitate. In a follow-up turn, "now add a unit test for the tie-break case" reuses the conversation's context instead of re-explaining.

4. Exam tips¶

🎯 Exam tip: know the difference between zero-shot (no example) and few-shot (one or more examples) prompting — and that few-shot is preferred when the output's format or style must match.

🎯 Exam tip: when a weak prompt yields a weak answer, the best fix is usually to add context or an example, or to reference the relevant file/selection — not to repeat the same prompt or switch models.

🎯 Exam tip: iteration and conversation history are features. Refining within the same chat thread is often the intended "best next step" because earlier turns remain in context.

5. Common pitfalls¶

⚠️ Pitfall: the "one perfect prompt" myth. Strong results come from specificity and iteration, not from a single magic sentence.

Vagueness: "make it better" — better how? Give a target.
No references: asking about "the config" without referencing the file forces a generic guess.
Overstuffing: a rambling, multi-goal prompt buries the intent and wastes the context window.
Ignoring history: starting a brand-new chat for a follow-up throws away useful context.
Wrong context open: unrelated open tabs can pull suggestions off-target — relevance beats volume.

6. Practice questions¶

1. Which prompt is most likely to produce correct, well-shaped code?

A. "Fix this."
B. "Improve the function."
C. "Refactor parseConfig to return a Result<Config, Error>, keep the signature, and add a test for a malformed file."
D. "Make it faster somehow."

Answer

Correct: C. It states intent, specifics, and a concrete edge case. A, B, and D are vague and give the model nothing precise to optimize for.

2. You need generated sample rows to match an exact JSON shape. Which technique fits best?

A. Zero-shot prompting
B. Few-shot prompting with two example rows
C. Lowering the temperature only
D. Opening more unrelated files

Answer

Correct: B. Few-shot examples pin the format. Zero-shot (A) leaves the shape to chance; C and D don't reliably control structure.

3. A first Copilot Chat answer is close but misses null handling. What is the best next step?

A. Start a new, unrelated chat and retype the whole request.
B. In the same thread, ask "now handle null and empty inputs, and add a test."
C. Switch to a different programming language.
D. Repeat the original prompt verbatim.

Answer

Correct: B. Iterating within the thread reuses conversation context. A discards context; C is irrelevant; D changes nothing.

4. In Copilot Chat, what is the purpose of referencing #selection or a specific #file?

A. It retrains the model on that file.
B. It explicitly adds that code to the prompt's context so the answer is grounded in it.
C. It disables content filtering for that file.
D. It increases the context window size.

Answer

Correct: B. Explicit references inject specific, relevant context. They don't retrain the model (A), affect filtering (C), or change the window size (D).

5. Which is the best description of prompt engineering for a developer using Copilot?

A. Writing code to fine-tune the model on your repo.
B. Crafting clear instructions and supplying the right context and examples to improve output.
C. Configuring the proxy server.
D. Training a new large language model.

Answer

Correct: B. Prompt engineering is about instructions, context, and examples — no model training or infrastructure work. A, C, and D describe unrelated technical activities.

Chapter 3 — Prompt Engineering & Context Crafting¶

In 30 seconds¶

Exam map¶

1. Key concepts¶

Anatomy of a strong prompt¶

Zero-shot vs few-shot¶

2. How it works¶

How Copilot determines context¶

Best practices that reliably help¶

3. In the real world¶

4. Exam tips¶

5. Common pitfalls¶

6. Practice questions¶

Further reading¶