Written August 2026 against Anthropic's current guides. Model names change fast, so the last section shows you how to check this yourself when they do.
Claude has 2 main models right now, and the little dropdown where you pick one doesn't explain itself. It reads like a good/better situation, where one is just the nicer version of the other.
It isn't. They're built for different jobs, and they want to be talked to differently.
If Opus has ever handed you 6 paragraphs when you wanted 1 line, that's not you prompting wrong. That's the model doing exactly what it does by default, and there's a single sentence that fixes it.
Which one to pick
Sonnet 5 for most of your day. Writing, planning, answering questions, going back and forth on an idea, cleaning up a caption before it goes out. It's fast, and it got a lot smarter this round.
Opus 5 when the job is big and you want it finished in one shot. A long document. A full landing page. Anything with a lot of moving parts that you don't want to babysit turn by turn.
Not sure? Start with Sonnet 5. Move up to Opus if the answer comes back thin.
1. Opus talks a lot. Sonnet matches your question.
Opus 5 runs long by default. You ask something small and get an essay back. Anthropic says the same thing in their own guide and tells you to just ask for shorter.
Sonnet 5 reads how big your question is and answers at that size. Small question, small answer.
So if you're on Opus and drowning in text, you have to say the word. It won't shrink on its own.
Paste this at the top of a chat, or into your project instructions so it sticks:
Keep responses focused, brief, and concise.
Keep disclaimers and caveats short. Spend most of the
response on the main answer.
When I ask you to explain something, give me a high-level
summary unless I specifically ask for the deep version.2. Delete "double check your work" from your Opus prompts.
This is the one that surprised me.
Opus 5 already checks its own work. Anthropic's guide says to remove verification lines from your prompts, because telling it to check again makes it do the entire job twice. You wait longer and you pay more for the same answer.
Same goes for "verify this before you respond" and "review everything at the end."
Sonnet 5 is the opposite, and this is the part people get backwards. Anthropic never tells you to strip those lines out of a Sonnet prompt, and Sonnet still takes direction well when you want it to go deeper. The nudge in section 3 below is a real one from their guide.
So it's not a rule you apply everywhere. It's an Opus move. If you've got a prompt sitting in your notes app from 6 months ago, go find that line and decide which model it's talking to.
3. Sonnet does exactly what you said. Opus fills in the gaps.
Sonnet 5 takes you literally. Anthropic puts it plainly: it won't assume an instruction meant for one thing applies to everything else. Say "fix the headline" and it fixes the headline, leaving the subhead and the button exactly where they were.
So spell out the scope instead:
Apply this to every section, not just the first one.Opus 5 goes the other direction. It makes a judgment call and will sometimes hand you more than you asked for. That's great when you want a partner on something big. It's annoying when you wanted 1 small edit.
To keep it in a narrow lane:
Deliver what I asked for, at the scope I asked for.
Make routine judgment calls yourself. Check in only when
2 different readings of my request would lead to very
different work.
If you think I'm wrong, say so in one sentence and keep
going with what I asked.And if Sonnet is being shallow on something genuinely hard, this nudges it deeper:
This task involves multistep reasoning. Think carefully
through the problem before responding.Say the whole thing in your first message
Both guides give this same advice, which almost never happens.
Most of us start small. You ask for one piece, see what comes back, then add the next bit. It feels careful.
Both models do worse that way. Anthropic says Opus 5 works best when you hand it the full job up front and let it run. The Sonnet 5 page goes further and says a task fed in over several messages gets you worse answers, not just slower ones.
So write the long messy version first. Then send it.
Here's the full job:
What I want: [the thing]
Who it's for: [the reader]
It has to include: [your must-haves]
Please don't: [what you don't want]
Done looks like: [how you'll know it's finished]It feels like too much. It isn't.
Go retest your voice profile
If you've built a voice profile or a "sound like me" prompt, run it again this week.
Anthropic's Sonnet 5 page says writing style can shift when a new model ships, and that if you count on a certain voice, you should re-check your style prompts against the new one.
Nothing is broken. Your profile didn't stop working. It was built against a model that writes a little differently now.
The test takes 5 minutes. Give it something you already have a good version of. Run it, then read the two side by side. If it drifted, add a line to your profile about what it got wrong. Same fix as always.
If you don't have a profile yet, that's the 7 steps.
The one trick worth stealing from the developer docs
Ask either model to design something with an open brief and it settles into a house style. You'll see the same look come back over and over.
Here's the trap: saying "don't use that color" or "make it cleaner" just moves it to a different fixed look. You get a new default, not variety.
What works, straight from Anthropic's Sonnet 5 guide:
Before you build anything, give me 4 different visual
directions for this brief.
For each one: background color, accent color, font, and
one line on why it fits.
Then stop and let me pick. Build only the one I choose.Use it on a landing page, a lead magnet, a carousel template. It's the difference between AI-looking design and something that looks like a person made a choice.
What you can skip
Both official guides spend most of their length on effort levels, token limits, thinking budgets, tokenizers, and error codes. All of that is for developers wiring Claude into their own software.
If you're typing into the Claude app, you can walk right past it. What's above is the part that actually reaches you.
Go check my work
Everything here came out of Anthropic's own prompting guides, so you can read the source yourself:
Prompting Claude Opus 5
Prompting Claude Sonnet 5
One thing to keep in mind, and it's the reason this page exists at all: these guides get rewritten every time a new model ships, which is often. A prompt you saved last spring was written for a model that doesn't behave that way anymore. Nothing was wrong with your prompt. It just expired.
Opening the source page takes 5 minutes, and it's the skill that keeps working long after this guide goes stale.