Anthropic released Claude Opus 5.5 on 22 September, and there's a page in their docs on how to prompt Claude Opus 5.5. It's written for developers, so most of it covers API settings an agency team will never see. I read the whole thing that week, and honestly, three bits of it are useful to anyone who types into Claude for client work. The first is a line you might be typing every day, and Anthropic's advice is to take it out.
The same page says prompts written for the last model, Opus 5, should still perform well without changes. So this is about getting more out of a model that Anthropic says writes more than 30 per cent faster than Opus 5 and tends to finish the same task with less output. I've tied each change to the agency jobs where I'd expect it to matter most: client briefs, monthly reports, proposals, and the spreadsheet with six tabs that only one person really understands.
Stop telling Claude to think carefully
If your team has a saved prompt, or a set of project instructions, that opens with "think carefully before you answer", you can probably take that line out. Anthropic's page covers this for chat apps. Its advice is to consider removing that kind of instruction, because Opus 5.5 works out for itself how much thinking a request needs. When Anthropic tested this in a chat product, replies started sooner with the line gone, and they saw no clear drop in the quality of the answers.
When a job needs more brainpower, change the effort level. Anthropic calls effort the main control for how much Opus 5.5 thinks. The default on Opus 5.5 is medium, and in Anthropic's testing, medium on Opus 5.5 matched or beat the old model running on high for coding and knowledge work. That's why I'd leave it at the default for most client work and stop fiddling with it.
I do most of my work in Claude Code, where there's an effort setting, and that's what I reach for now when a job is hard. I'd push it up a level for the work where a weak answer costs you, like the pitch strategy or the reply to a client who's unhappy. Anthropic's advice is to keep the two highest settings (called xhigh and max on the developer side) for work where you've measured a real gain. So try one level up on one job and compare the two answers before you make it the team habit.
Tell it to look around before it starts
Opus 5.5 gets stuck in fast (Anthropic's phrase is that it tends to get to work quickly), which is mostly good news, but it can miss things your request didn't mention. Their examples will sound familiar to anyone who runs accounts: a policy in an old email thread, a rule on another spreadsheet tab, a note on a customer record.
Say the brief asks Claude to refresh the summer campaign, and the client's rule about never showing prices is in an email from March. Or the monthly report pulls from the main tab of the spreadsheet, and the note saying one campaign was paused for a week sits on a tab called "notes". Claude does the job you described well, and you still can't send it. That sounds obvious, but it's one of the hardest mistakes to catch, because the draft reads well.
It's worth fixing, because once it's looking at the right material it's sharp. Anthropic says it catches details that are easy to miss in big inputs, like a date in a long planning thread that falls on the wrong weekday, or a chart in a slide deck that doesn't match the figures underneath.
The fix is one sentence. Anthropic's version is written for agents with tool access, so this is how I'd put it for someone working in the Claude app:
"Before you do anything, look through every email, doc and spreadsheet tab that could matter to this task, including ones I haven't mentioned, and use what you find."
In Anthropic's testing on tasks that ran across several connected apps, Opus 5.5 got noticeably more of them right with that sentence in place. They don't give a number, so I won't invent one. It costs a few extra steps and a bit more usage, which is a fair trade on anything going to a client.
It can only look at what it can reach, though, so this matters most when Claude is connected to your email and files, or when you've attached everything relevant. The page also carries a warning: because the sentence tells it to act on what it finds, keep untrusted content out of what it searches. For an agency, I'd read that as being careful before you point it at a shared inbox full of emails from people you don't know.
Brief it on style by naming what you don't want
This one comes from the page's section on web design, and I'd use it for anything with a look or a voice. Anthropic found that a general line like "avoid a generic AI look" mostly trades one default style for another, so you get a different kind of bland.
Anthropic's fix is to name the specific things. Their example asks for a personal website with no cream or off-white background, no italic accent words in headlines, no numbered "01/02/03" section labels, no monospace labels and no pill-shaped buttons. Every item on that list is something you can see and check. Then the advice is to look at what it used instead and add that to the list next time.
It's the same brief you'd give a new copywriter in their first week: here are the things we never do. For a client report, I'd name an opening line about the fast-moving world of whatever, a bold label at the start of every bullet, a final paragraph that repeats the summary, and dashes used where a comma would do. For a proposal deck, name the layout habits you're sick of, such as a question as the title of every slide.
I keep a list like this for my own writing, and it gets longer every time I catch something new. Your team's first list will be short, which is fine. Every time someone spots a pattern they hate in a draft, it goes on the list, and the list lives somewhere everyone pulls from.
Getting your whole team to prompt Claude Opus 5.5 this way
None of these takes more than a minute for one person. The hard part, in a team of 40, is getting everyone to make the same change, including the people who never read release notes.
So put the lines where the work starts. If your team keeps a shared doc of Claude prompts, or uses shared projects in Claude with their own instructions, the look-around sentence and the style list go in there, and the "think carefully" line comes out. Then someone checks a fortnight later that people are using them. I wrote about why getting a process written down matters in this post on SOPs, and about why a manager backing the change makes the biggest difference in this one on adoption.
If you only change one thing this week, add the look-around sentence. It's aimed at the mistake most likely to reach a client: a draft that reads well and misses the rule sitting in last March's email.
Getting a whole agency team working like this on real client work is what I do, and you can see how that works here.