Anthropic Revealed Their Secret Guide to Mastering Opus 5.5

0:00 Enthropic just dropped this guide on how to use Opus 5.5. And if you want to get the most out of it, I'm going to speedrun eight tips covering the key takeaways, what's different, what to

0:09 change, and practical examples you can use. And a few of these I had absolutely no idea about. This should not only help you get better results with this model, but also save you money on token costs.

0:20 So, without further ado, let's dive right in. The first concrete change is the default effort setting. Opus 5 defaulted to the high effort level while Opus 5.5 now defaults to medium.

0:31 Basically, effort controls how much work the model puts into thinking before it responds. And Enthropic recommends starting at medium and testing it on your own tasks. And don't assume that

0:41 you need to carry over the higher setting that you used with the previous models. The same effort level doesn't necessarily mean the same amount of thinking across these models. If you

0:50 keep your old configuration, you could get longer turns and more output tokens. Enthropic recommends lowering the effort level first when you want less thinking and then only using the highest settings

1:00 where you've actually measured a benefit. This way, it's going to save you a lot on token usage. Otherwise, you're just going to use a higher model even though it might not be necessary

1:09 with Opus 5.5. And then for developers, the guide also covers leaving enough room for thinking tokens and changing effort without invalidating the prompt cache. But if you're not a developer,

1:19 this probably sounds like gibberish, so it probably doesn't apply to you. A useful way to choose your setting is to compare the same task at different effort levels. For example, you can give

1:28 Claude two proposals in the same list of requirements, then compare both medium and the higher effort. And then you could ask yourself, did one miss an important requirement? Was a

1:37 recommendation more useful? How long did it take? A longer answer alone doesn't necessarily tell you whether the effort level was worthwhile. And this is going to be custom to you and your individual

1:47 use cases. All right, moving on to tip number two, and this is thinking instructions. This is worth checking if you've built up a longer system prompt over time inside of Claude. Enthropic

1:57 says that instructions like thinking carefully before answering might be unnecessary for Opus 5.5 because the model already decides how much to think and the effort is the main control. In

2:08 the specific test that Enthropic ran, removing those kinds of instructions made replies start sooner, and this meant you're getting quicker replies, and it even did so with no clear decline

2:18 in the quality of the reply. Another behavior that Enthropic documents is that Opus 5.5 can go back over an earlier answer while thinking about a new message, even a simple follow-up.

2:29 The guide gives an instruction that tells it to treat previous answers as settled unless you question them. that can reduce unnecessary work. But there is a trade-off. You may want Claude to

2:39 revisit earlier conclusions during research, analysis, or a task where new evidence appears. Now, right here is a perfect example. When you write out a vague prompt, something like thinking

2:49 deeply and carefully, it doesn't necessarily tell Claude what makes that answer useful to you, and that's probably one of the reasons you're not getting the outputs that you're wanting.

2:58 Whereas, if you give it a specific prompt, like asking it to compare two proposals against your requirements, recommend one, and explain the made trade-off, gives it a concrete target to

3:08 work towards. You're basically still asking for thoughtful work, but you're defining the result that you need. Next up is tip number three, and this is relevant context. Enthropic says that

3:17 Opus 5.5 tends to get to work very quickly. This can be very useful, but in a workflow involving several apps, the task might depend on information that the request doesn't explicitly mention.

3:28 This could be something like a deadline that could have changed in an email or a rule that could be on another spreadsheet tab, for example. Their recommendation is to ask Claude to look

3:37 through relevant sources before acting rather than assuming the first document contains everything. This way, it has all of that relevant context. This screenshot right here shows the actual

3:47 instruction that Enthropic provides. It tells the agent to explore the emails, documents, spreadsheet tabs, and records that could be relevant, including ones the task didn't specifically name.

3:58 Enthropic found that this improved completion on its own multi-app test with slightly more tool calls and tokens. The important conditions are that the sources are accessible and that

4:08 the agent isn't necessarily blindly acting on untrusted material. Now, let's move on to an example. Imagine you ask Claude to write a project update from an original brief. And let's say that this

4:19 brief says the deadline is Monday, but a later email moves it to Friday. If Claude only reads the brief, it can produce a polished update with the wrong date. Obviously, this prompt that you

4:29 see right here asks it to check both sources before drafting. So, this way, you're giving it a specific reason to look beyond that first file that we gave it. Moving on to tip number four, and

4:39 this is pasted material. When you paste an email or web page into a message, there are really two different things present. your request and somebody else's content. That content might

4:49 include instructions, but those instructions aren't necessarily yours. Enthropic recommends marking the boundary clearly. The everyday principle is simple. Make it obvious what you want

4:59 Claude to do and which material you're asking it to read or analyze. So, just start with a job basically like saying summarize this email. Then, paste the email in separately so the request and

5:08 the source material are clearly separated. This is very simple but just really important when prompting Opus 5.5. All right, so moving on to tip number five. Why Claude can seem silent.

5:18 Opus 5.5 can produce progress updates that a custom app does not display. If you built the app, check that it can receive and show those updates. Asking Claude to talk more will not fix updates

5:29 the app is hiding. You can tell Claude when you want an update rather than leaving the timing open. Just make sure to choose useful moments when you're starting the work, finishing the review,

5:37 and delivering that result. for automated setups. The guide also explains how to send a limited number of reminders. All right, so here is another example. Here we ask for an update after

5:47 the files are reviewed in when the draft is ready to send. Claude should keep working between these different updates. This gives you visibility without turning every step into an approval

5:56 request. Now, moving on to tip number six. A finished reply is not a finished task. With Opus 5.5, a progress update can end a turn while work is still outstanding. List what you expect to

6:08 receive and keep track of the unfinished items. For an automated agent, the app must also check for completion. And a checklist alone is not a guarantee. Now, if a background task is still running,

6:18 wait for its results before calling the job done. Anthropic provides extra instructions for agents that run without someone watching. Those instructions can use more time and tokens, and they

6:28 should still preserve necessary approvals. This longer instruction that you see targets a simple problem, reporting the next step instead of taking it. It asks an unattended agent

6:37 to continue work that does not need your input. Real blockers and required approvals still matter. Now, right here is an example. We need to tell Claude what done includes first. We can name

6:46 the outputs, whether that's a report, source list, and a summary. If one of those is missing, point to it directly instead of just saying keep going. Ask Claude to finish the item or just

6:55 explain the blocker. Moving on to tip number seven, and this is giving Claude a clear design direction. Opus 5.5 still falls back on default website styles. Anthropic says less generic often swaps

7:08 out one default for another. So give concrete visual feedback and refine what comes back. This is obviously just an existing best practice and not a new feature baked into Opus 5.5, but I do

7:18 think it is relevant for you to understand when prompting this model. Now moving on to a very simple example. We need to be specific about the design. And so this is a perfect example of

7:26 this. When we say make it less generic, this leaves most of the choices open and it's going to default to the next design style. The stronger example names a background, headline style, button

7:36 shape, and specific spacing. We can use our own preferences or a reference image as an example. Moving on to tip number eight, making small details easier to read. Anthropic reports that Opus 5.5

7:47 reads visual material more accurately than Opus 5 without extra tools. Dense charts and tiny labels can still benefit from a sharp original and a close-up. If image tools are available, Claude can

7:58 crop and inspect the details itself. Now, here's a quick example. Point Claude to the exact label or part of the image that matters. Let it crop if tools are available or attach your own

8:07 close-up alongside that original image. Then, you could ask it to say when the detail is unreadable rather than guess, which is going to help provide more context here. All right, so I know we

8:16 talked about a lot of different tips here, and this might be overwhelming. You don't need to add every single recommendation to one enormous prompt or you don't need to test out every single

8:24 one of these. But hopefully this helps you prompt Opus 5.5 in order to get better responses and also save you money on token usage since that is obviously something that you're wanting to do. If

8:34 you enjoyed this video, subscribe to the channel for more content like this. Let me know what you think in the comments and I'll see you in the next video. Cheers.