Lesson 2: Giving Rich Context and Examples

Why Context Is the Biggest Lever

In Level 1 you saw that people who get more from AI give it more context. This level shows you how to do that on purpose.

Anthropic's prompting guide for Claude describes the assistant as a brilliant but very new employee. It is capable, but it knows nothing about your organisation, your readers or what you count as good work. The guide says results improve when you explain what the output will be used for, who it is for, what larger workflow it belongs to and what success looks like. OpenAI and Google give similar advice: be specific, give background, and show examples of what you want.

An assistant can't ask what you meant before it starts. Anything you leave out, it fills with a generic guess.

Task 1: Rewrite a Vague Prompt with the Context Checklist

Use these four questions on every prompt that matters:

• Audience: Who will read or use the output? What do they already know, and what do they care about?
• Purpose: Why does this task exist? What should happen after someone reads the output?
• Background: What facts does the AI have no way of knowing? For example, earlier decisions, constraints, history or sensitivities.
• Success criteria: How will you judge the result? Write two or three checks you could tick off.

Before:

Write an email about the project delay.

After:

I need an email to our client, the operations director at a mid-sized logistics firm. She is not technical and cares about her go-live date. Purpose: tell her the warehouse system launch is moving from 3 March to 17 March, and keep her trust. Background: the delay comes from a supplier's late hardware delivery, not from our team. She was already unhappy about a two-day slip last month. We are offering free on-site support during the first launch week. Success criteria: (1) the new date appears in the first two sentences; (2) the email states the cause without blaming anyone by name; (3) the support offer is concrete; (4) it is under 150 words.

The second prompt takes about a minute longer to write. In return, the AI no longer has to guess the tone, how much to explain or what to put first.

Do this now:

1. Pick a real prompt you used recently that gave a disappointing result.
2. Write one line for each of the four checklist items.
3. Write the success criteria as checks you can test, such as "new date in the first two sentences". "Make it good" is not a check.
4. Save both versions. You will need them in Task 3.

Task 2: Add Examples the AI Can Follow

Some things are hard to describe but easy to show, such as a house style, a level of detail or a way of structuring a summary. In those cases, an example works better than a description.

1. Start with one to three examples. OpenAI's guidance suggests trying without examples first, then adding a few if the output misses. Anthropic's Claude guide suggests 3–5 examples when you need a consistent structure. Start small and add more only if the output is still inconsistent.
2. Label them clearly. For example: "Here are two past updates I liked. Match their tone and length, not their content." Anthropic suggests putting each example inside tags such as <example> so the AI can tell where it starts and ends.
3. Make them relevant. Examples should be close to your real task. A cheerful marketing post is a poor guide for a formal incident report.
4. Make them different from each other. If all your examples open the same way or use the same topic, the AI may copy that pattern. Anthropic's guide warns that examples should vary enough that the model does not pick up unintended patterns.

How to spot over-copying:

• The output reuses phrases, sentence openings or facts from your example.
• Every output has exactly the same length or structure, even when the content needs something different.
• Details from the example, such as names or figures, appear where they don't belong.

Fixes: add a second example that differs in surface details, and say directly what to keep ("match the tone and structure") and what to ignore ("do not reuse any wording or facts").

Task 3: Test Before and After Against Your Criteria

After writing a better prompt, check whether it actually worked and find out why.

1. Run your original vague prompt. Save the output.
2. Score it against your success criteria: one tick per criterion met.
3. Add context one piece at a time: audience, then purpose, then background, then examples. Run the prompt again after each addition. Use a new chat each time so earlier answers don't affect the result.
4. Score each version against the same criteria.
5. Note which addition made the biggest jump. That tells you what this kind of task needs most.

You will usually see a pattern. For client emails, the background often matters most. For summaries, it is often the audience. For anything in a house style, it is often the example. Keep a short note of what you find, because later levels build on it.

Self-Check

• Can you name the four checklist items without looking?
• Does your rewritten prompt contain at least one fact the AI could not have guessed?
• Can each success criterion be clearly passed or failed?
• Did you check the output for copied wording from your examples?

Key Takeaways

• Treat the AI like a capable new colleague who knows nothing about your situation.
• Use the checklist: audience, purpose, background, success criteria.
• Write success criteria you can test, so you can judge results objectively.
• Start with one to three relevant, varied, clearly labelled examples.
• Watch for over-copying, and say what to keep and what to ignore.
• Add context one piece at a time to learn which piece improves the result most.