Instructions vs Context vs Examples: What Is Your Prompt Missing?

If you’ve used Claude, ChatGPT, or another AI assistant, you’ve probably had this experience: the prompt seems perfectly reasonable, but the answer still isn’t quite right.

Maybe it answers the general topic without doing what you wanted. Maybe it does the right task but gives you something so generic that it could apply to anyone. Or perhaps the information is right, but the tone and structure are completely wrong.

The temptation is to make the prompt longer.

That isn’t always the answer.

A weak result can come from different gaps in the prompt, and adding more detail in the wrong place may do very little to fix it. The useful question is not simply “How can I write a better prompt?” but “What is this prompt missing?”

Three things are particularly useful to check: instructions, context, and examples. Each solves a different problem.

If you’re new to prompting and want to learn how to build a strong prompt from the beginning, start with our How to Get Better Results from Claude guide. Here, we’re going to look at what to do when you’ve already prompted an AI and the result isn’t working.

Why This Distinction Matters

Suppose you ask an AI to help with your pricing page and dislike the response.

There are several possible reasons.

Perhaps you never told it what you wanted done to the page. That is an instruction problem.

Perhaps you told it exactly what to do, but never explained who your customers are, what you sell, or why the current page isn’t working. That is a context problem.

Or perhaps the task and background were clear, but the copy sounds nothing like your brand. An example might be what is missing.

Those problems can feel similar when you are looking at a disappointing response, but they need different fixes.

Treating every weak response as a sign that “the prompt needs more detail” can leave you with increasingly long prompts that still don’t produce what you want.

A better approach is to diagnose the gap first.

Instructions: When the AI Does the Wrong Thing

Instructions define the job.

The key question is:

What should the AI do?

If the response talks about the right subject but doesn’t perform the task you had in mind, your instructions are the first place to look.

For example:

Help with our pricing page.

There is nothing technically wrong with that request, but “help” could mean almost anything. The AI might critique the page, suggest a new structure, rewrite the headline, discuss pricing strategy, or give you general conversion advice.

Now compare it with:

Rewrite this pricing page section to lead with value instead of features. Keep it under 100 words.

The subject hasn’t changed. The job has.

Good instructions can tell the AI what action to take and, where necessary, set boundaries around the result: the format, length, scope, audience, or objective.

A useful test is:

Could two people reasonably interpret my request as two different tasks?

If yes, the instructions probably need work.

Signs you may have an instruction problem

The response:

• discusses the topic instead of completing the task
• gives you the wrong type of output
• covers far more than you wanted
• focuses on the wrong part of the problem
• ignores a format or length requirement you never stated

The fix is not necessarily a longer prompt. It may simply be a clearer verb.

“Help with X” gives the AI a subject.

“Compare X,” “rewrite X,” “summarise X,” “find the weaknesses in X,” or “turn X into a table” gives it a job.

Context: When the Answer Is Right but Could Apply to Anyone

Context is different. The AI may understand the task perfectly and still lack the information required to tailor the answer to you.

The key question is:

What does the AI need to know about my situation for the answer to be useful?

Imagine asking:

What should I consider before increasing my prices?

The instruction is clear. You want factors to consider before a price increase.

But the AI knows almost nothing about the situation.

Now add:

We’re a two-person consulting business selling mainly to enterprise clients. Our sales cycle is around six months, and most clients stay with us for several years. We’re considering increasing our prices for new clients.

The task hasn’t changed, but the information surrounding it has.

That context could affect which risks, trade-offs, and considerations deserve the most attention.

Useful context might include your business model, audience, objective, constraints, previous attempts, resources, or decisions that have already been made.

You do not need to tell the AI everything you know.

A good test is:

If I removed this detail, could a sensible recommendation change?

If yes, it is probably useful context.

If not, you may simply be adding noise.

Signs you may have a context problem

The response:

• sounds generic
• makes assumptions that do not fit your situation
• recommends things you have already tried
• ignores an important constraint
• gives technically sensible advice that would be difficult for you to use

This is where people often make prompts longer in the wrong way. They add more instructions when what the AI really needs is more information about the situation.

Examples: When the Answer Is Right but Doesn’t Feel Right

Sometimes the AI understands the task and has enough context, but the finished result still isn’t what you pictured.

This is particularly common with subjective qualities such as:

“Professional.”

“Warm.”

“Punchy.”

“Premium.”

“Conversational.”

“Not too salesy.”

Those descriptions are useful, but they leave room for interpretation.

That is where examples become powerful.

The key question is:

What should the finished result look or sound like?

Instead of saying:

Make this professional but approachable.

you could provide a short piece of copy that already has the tone you want and say:

Match the tone and level of formality of this example, without copying its wording.

An example gives the AI a target rather than asking it to interpret a collection of adjectives.

Examples can be particularly useful for:

• tone of voice
• headlines
• email structure
• product descriptions
• reports
• formatting
• recurring brand content

You might provide a sentence written in the right tone, a previous email you liked, an example of the structure you want, or an existing piece of your own content that represents the brand well.

You don’t need examples for every prompt. They become valuable when describing the target is harder than showing it.

The Diagnostic Framework

When a response isn’t working, start with the result rather than immediately rewriting the entire prompt.

What went wrong?	Check first	What to change
It talked about the topic but didn’t complete the task	Instructions	Make the required action explicit
It gave the wrong type of output	Instructions	Specify the format or deliverable
It covered too much or focused on the wrong area	Instructions	Narrow the scope
It did the task but the answer was generic	Context	Add information specific to your situation
It made an assumption that doesn’t fit your situation	Context	State the missing fact or constraint
The content is useful but the tone feels wrong	Example	Show a representative sample
The structure isn’t what you pictured	Example or instructions	Show the target structure or specify it directly

This is not a rigid formula. A response can have more than one problem.

But you usually do not need to fix everything at once. Identify the most obvious mismatch, correct it, and see what Claude does next.

One Prompt Can Have All Three Problems

Consider this:

Write something for our new service.

The AI has very little to work with.

The instruction is unclear: what are you asking it to write?

The context is missing: what is the service, who is it for, and why should anyone care?

And there is no example to guide the style if the tone matters.

You could turn it into:

Write a 120-word announcement email introducing our new monthly bookkeeping service to existing freelance clients. The service gives clients a monthly review of income, expenses, and upcoming tax obligations so they have a clearer picture of their finances throughout the year. Keep the tone straightforward and reassuring rather than promotional.

For tone, aim for something similar to: “A quick update from us — we’ve added a new way to make staying on top of your finances a little easier.”

Now each part has a job.

Instruction: write a 120-word announcement email.

Context: existing freelance clients, monthly bookkeeping service, what it does and why it matters.

Example: the sample sentence establishes the intended tone.

That is the point of separating the three ingredients. You can see what each piece contributes instead of thinking of a prompt as one big block of text.

The Most Common Misdiagnosis: Adding More Instructions When You Need Context

This is probably the mistake worth remembering.

When an answer feels generic, it is easy to assume you need a more elaborate prompt. So you add more rules:

Make it detailed. Be specific. Give practical recommendations. Don’t be generic. Make the advice actionable.

But none of those instructions tells the AI anything new about your situation.

If it does not know your budget, audience, business model, constraints, previous attempts, or whatever else materially affects the answer, telling it to “be more specific” can only go so far.

For example:

Give me specific and actionable ideas for marketing my business.

still leaves the AI guessing.

Compare that with:

I run a two-person bookkeeping firm that gets most new clients through referrals. We have no dedicated marketer and around five hours a week available for marketing. Give me three channels worth testing before we consider paid advertising.

The second prompt doesn’t merely demand specificity. It provides the information required to be specific.

That is the difference between adding instructions and adding context.

Don’t Add an Example When an Instruction Would Do

The reverse can happen too.

Suppose you want five subject lines under 40 characters.

You could spend time finding examples of short subject lines you like, but if style is not particularly important, you probably don’t need them.

Just say:

Give me five email subject lines, each under 40 characters.

Examples are useful when they communicate something that would otherwise be difficult to explain. They are not an ingredient you have to force into every prompt.

The same applies to context. You don’t need to explain your entire business before asking Claude to correct a spelling mistake.

The goal is not to build the longest possible prompt.

It is to give the AI the information the task requires and no more than is useful.

Practical Takeaway

When an AI response falls short, don’t immediately rewrite the whole prompt.

Diagnose the mismatch first.

Did it do the wrong thing?
Check the instructions.

Did it do the right thing, but give you an answer that could apply to almost anyone?
Check the context.

Did it understand the task and situation, but produce something that doesn’t look or sound right?
Consider an example.

Sometimes you will need two or all three. Sometimes one clear sentence will fix the problem.

The point isn’t to make every prompt longer. It is to give the AI the right kind of information at the right time.

FAQ

Do I need all three ingredients in every prompt?

No. A simple factual request may only need a clear instruction. Context becomes more important when the answer depends on your particular situation, while examples are especially useful when style, tone, structure, or format matters.

Which ingredient matters most?

There isn’t one ingredient that matters most in every situation. Instructions define the task, context makes the answer relevant to your circumstances, and examples can clarify what the finished result should look or sound like.

Can too much context hurt a prompt?

Irrelevant context can make a prompt unnecessarily complicated without improving the answer. Focus on information that could meaningfully affect what a good response looks like.

How many examples should I give?

For many everyday tasks, one representative example may be enough to communicate the target. More complex or highly consistent work can benefit from several examples, particularly when you want the AI to recognise a recurring pattern.

What if I’m not sure which ingredient is missing?

Start with the response you received. Identify the most obvious mismatch: the task, the relevance, or the result itself. Fix that first and try again rather than changing several things at once.

Is this only useful for Claude?

No. The distinction between instructions, context, and examples is useful when working with other generative AI assistants as well.