A lot of people working with Claude for the first time do the same thing: they type a question the way they would ask Google, get a mediocre answer, and decide it is overrated. It is rarely the model’s fault. It is almost always the instruction.
A prompt is not a search query. It is closer to a brief, the kind you would give a new intern. A good intern who knows nothing about your situation can do excellent work, but only if you say what you want, why, and in what form. Give too little and they have to guess. And since Sonnet 5, the model guesses less. That is exactly why this is a good moment to start.
Why “more literal” is good news
Sonnet 5 follows your instructions more precisely than earlier versions. It fills in fewer gaps. For an experienced user that takes some getting used to, because the vague prompts that used to work anyway now work less well.
For a beginner it is a gift. You do not have to learn tricks to make the model guess your intent. You say what you want, and you get it. If you get something other than what you had in mind, you know straight away where it went wrong: your instruction was not clear enough. That is a mistake you can fix yourself, and you get better at it every time.
The three parts of a good instruction
Almost every good prompt has three things in it: a goal, context, and a form.
The goal is what you want to happen. Not “something about quotes”, but “write a quote”. Be concrete about the task itself.
The context is everything the model needs to know to do it well. Who is it for? What is the situation? What absolutely must or must not be in it? This is the part beginners skip most often, and it is usually the difference between a usable answer and a useless one.
The form is what the answer should look like. A list or flowing text? Long or short? Formal or informal? Since Sonnet 5 decides the length itself when you say nothing, name your preference if you have one.
Wrong versus right
A few examples make it concrete.
Wrong: “Write an email to a customer.”
This could go anywhere. Which customer, about what, in what tone? You get something generic back that you have to rewrite from scratch anyway.
Right: “Write a short, friendly email to a customer who has not yet paid their invoice. It is two weeks past the due date. I do not want to come across as harsh, it is a good customer. Ask whether something is the matter and gently remind them of the outstanding amount. No more than eight sentences, informal tone.”
The second example has a goal (write an email), context (good customer, two weeks late, not harsh), and a form (short, informal). You get something back you can send almost as is.
One more:
Wrong: “Help me with my planning.”
Right: “I am a freelancer and I have five jobs and two deadlines this week. Help me build a daily schedule for Monday to Friday. I work best in the morning, so put the heavy work there. Give it back as a schedule per day.”
Five beginner prompts that always work
You can use these five today. Adjust the details to your own situation.
1. Explain. “Explain [topic] as if I know nothing about it. Use plain language and an example from everyday life.”
2. Make it shorter. “Cut this text in half without losing the core. Keep the same tone.” (Then paste your text.)
3. Improve. “Check this text for spelling and clumsy sentences. Do not change the content or my tone, only fix what is genuinely wrong or unclear.”
4. Brainstorm. “Give me ten ideas for [topic]. Keep them short, one line per idea. Wild ideas are welcome, I will filter afterwards.”
5. Push back. “I am considering [decision]. Play devil’s advocate for a moment: what are the reasons not to do it?”
That last one might be the most underrated. Many people use Claude only to confirm something. Asking for pushback often yields more.
One thing to keep in the back of your mind
The model guesses less than it used to, but it does not check whether something is true. It sounds convincing even when it is off. So always read the answer yourself before you use it, especially with facts, figures and names. And watch what you paste in: do not drop customer data or confidential information into a chat without knowing where it ends up. The simplest rule is that you stay responsible for what you send, not the tool.
The next step
Start small. Take a real task from today, write a prompt with a goal, context and form, and see what comes out. Not good enough? Adjust your instruction instead of leaving it at one attempt. That back and forth is not failure, it is exactly how you learn.
Once this starts to click, you will probably want more. In the next part I go deeper into the controls Sonnet 5 gives you: how hard it works, how long it answers, and how to make it check its own work.
Read on in part 3: getting more out of Sonnet 5.