Once you have the basics down, the difference between a decent answer and a good one usually is not a better prompt sentence. It is the controls around it. Sonnet 5 gives you a few settings most people never touch, and that is exactly where the gains are. This part is about those controls.
Effort levels: how hard the model works
Sonnet 5 has five levels for how much effort it puts into a task: low, medium, high, xhigh and max. High is the default. The idea is that you choose per task how much compute and time you want to spend.
In practice it comes down to this. Low is for simple, repetitive work where you want speed: formatting a list, rewriting a short text, converting a file. Medium to high is your daily workhorse, good for almost everything. Xhigh and max you save for the real head-scratchers, a tricky piece of code or an analysis where you cannot afford a reasoning slip and that is allowed to take a while.
A handy reference point: Sonnet 5 at medium performs roughly like the previous version at high, and Sonnet 5 at high like the previous version at its very top. So you get more reasoning depth at the same or a lower setting than you were used to. Do not put everything on max by default. It costs time and money without giving you anything back on most tasks.
Steering the length
Sonnet 5 decides the length of an answer itself. A simple question gets a short answer, an open-ended one gets more. Usually that judgment is right, and then you do not have to do anything.
But sometimes what you need differs from what the model expects. Then you have to say so, and say it concretely. “Be concise” is vague. “Answer in no more than three sentences” or “give me a full write-up of at least 800 words” is clear. Because Sonnet 5 follows you more literally, a concrete instruction like that actually works. Ask for three sentences and you get three sentences.
A common mistake: people complain that answers are too long or too short, but never say what they do want. Name it. The length is in your hands now, so use it.
Verification loops: have it check its own work
This is the technique with the biggest effect and the least use. The idea: give Claude a way to check whether its own output is right, and the quality goes up noticeably.
An example. Ask for a calculation, then add: “After that, check your answer by reversing the sum and seeing whether you arrive back at the starting numbers.” Or for a text: “Go through every claim in this piece and check it against the source I gave you, and flag what you cannot verify.” For code: “First write a few tests that would fail if the solution is wrong, run them, and adjust your code until they pass.”
What you are really doing is giving the model a yardstick to measure itself against. Without that yardstick it produces something that sounds plausible. With it, it produces something it has actually tested. The difference is large, and it costs you one extra sentence in your prompt.
Context management: stop repeating yourself
If you notice that you explain the same background in every conversation, you are wasting time. Who you are, how you write, which rules apply, the model only needs to know that once.
In Claude Code you do this with a file called CLAUDE.md: you put your context, preferences and rules in it, and the model reads it automatically with every task. In the regular Claude app you can achieve the same with project instructions or a fixed block of text you paste up front. The principle is the same: write the context down once, in a place that travels with you, instead of repeating it every time.
Treat such a file as a living document. Notice the model making the same mistake or ignoring a preference of yours? Put the correction in it. Next time it already knows. After a few weeks you have an instruction that knows how you work, and that saves you in every conversation.
Updating your old prompts
Your prompts from the previous version still work, but a few things you can now drop or sharpen.
Drop the manual thinking instructions. Did you explicitly ask the model to “think carefully first” via the old extended-thinking setting? That is no longer needed and via the API it even returns an error. It thinks for itself now.
Remove the temperature settings. Did you work via the API with settings like temperature or top_p to make the output more or less predictable? Sonnet 5 no longer accepts them and returns an error if you send them along.
Make your vague prompts more concrete. Where you used to lean on the model’s ability to fill in the blanks, you are better off asking explicitly now. It guesses less, so say more.
And if you work via the API: keep in mind that the same text is now counted at roughly thirty percent more tokens by a new tokenizer. That affects your costs and how much you pass in at once. More on that in the next part.
What it costs, and not only in money
A high effort level makes the model compute longer and harder. That costs more money, and it costs more energy too. For a single question you will not notice, but if you put everything on the heaviest level by default, it adds up, both on your bill and in the power behind it. The same thought applies to the new tokenizer: the same text now counts heavier, so a habit like “I will just throw the whole folder in to be safe” is more expensive and more wasteful than it looks. The middle level is not only cheaper, it is also the leaner choice. Use the heavy level where it is genuinely needed, not out of convenience.
The next step
With effort levels, steering the length and verification loops you already get noticeably more out of it. With a good context file you stop repeating yourself. For most advanced users that is plenty.
Want to go all the way to the bottom? In the final part I move to the power-user approach, with the tips from the maker of Claude Code, parallel sessions, and how to choose between Sonnet 5 and the heaviest model.
Read on in part 4: expert prompting with Claude Sonnet 5.