Opus 5 Is Exhausting. Anthropic Reveals The Fix. transcript

Okay, so one of the biggest problems that you've likely been facing with Opus 5 recently is the fact it's really confusing when outputting any text. So for example, over here I was chatting about some ideas and it basically said, "The corpus is an engine. Angles come from what you've already ingested." And I read through a lot of this output and I'm kind of like, what on earth does this even mean? I have no idea. It feels like my English has become weaker or it's speaking another language.

And I know I'm not the only one feeling this way because there's a pretty recent viral blog post where basically the author says, "Reading AI output is extra effort. It's verbose, frequently contains all too plausible nonsense, and it is increasingly jargon-dense." And recently they got this output from Claude, which I would not even try reading. And it said, "I had to look up every single word to make sense of this." And basically the solution that someone from the Claude Code team recommended is using output styles.

And I actually did make a video about this inside of my agentic coding school. Almost like 8, 9 months ago now, back in October when it was released, but I haven't really used it since. So if I go over to output styles over here, I basically start out the video by saying that this is a feature that I don't use much, but it does exist. But I have been using it more and more because this helps get around the fact that Opus 5 is really annoying when it comes to outputting.

So I'm going to quickly go through how we can set this up over here. So this will be down below so you can literally just copy the text. But she essentially says that she likes to use this after a long day, Honestly helps so much. And it seems that some of the team members like using different output styles for either different projects, different tasks, or even at different times of the day, depending on how tired they're feeling or how engaged in the process they want to be.

So it's not a case of you make one output style, then you set it once. You will be switching back and forth between them over time. So anyways, I'll quickly go through how you can set this up. Basically, this will be down below, so you can literally just copy the text. Then go over to Claude Code, tag the Claude Code guide by doing @claudecodeguide, paste in the output style and say, and say, add this output style for me, press enter.

And then it will go ahead and do that. And then once it is done, you can do /config inside of Claude Code, search for output style over here, and then switch over to Explain Like I'm 5, and then press escape. And then you can just rewind to go back earlier one message, and then basically enter in the same prompt again. And then after entering the same prompt again, like the Explain Like I'm 5 output style, it made much more sense.

Now you will have noticed there are other output styles as well. So there is a learning output style where it will ask you to write some code yourself. There is also the explanatory output style, which I know the Claude Code team likes to use because they mentioned it in an interview where they get new hires to basically enable that when they're working inside of a brand new repo. Whenever any new engineer joins our team, we told them to use the exploratory output style.

So this is just like /config output style equals exploratory. You run this in Claude Code, or you can ask Claude to set this for you. And what it does is anytime Claude will make a change, it'll explain to you, hey, here's how the architecture works. Here's how this language works if you haven't used it before. Here's how this part of the codebase works. It'll explain to you so that you can learn. Now, of course, this is not a one-size-fits-all output style.

You may want to kind of experiment around and find a better one for your liking. And one way that you can do this is that anytime you come across some really confusing or annoying output, kind of like I have in this chat over here, you can basically do /branch and then it will branch from the existing conversation onto a brand new chat. And then I can basically paste in the existing output style and say, hey, so can you basically generate me like 5 other output styles and how your previous output would look like in that brand new style?

Essentially my goal here is to find a style of communication which I understand. And I can see there's a bunch of output styles. So there is a kid mode, which I will not be using. That's way too basic. There's also the Slack DM over here, which looks a bit better. But honestly, when I'm reading for it, then it's like still pretty confusing. And if you find that doesn't work for you, then you can ask it to make an output style based on ASD-STE100, which is a language standard, which is for Simplified Technical English.

And if I were to basically rewind and then say like, okay, can you explain, explain in this instead, then it will give a much better style. And you may want to adjust this slightly. And if I were to make this into a brand new output style, then I may find over time, actually I want it to be a bit more technical. And now reading through this, I can be like, okay, wow, this is significantly better. I want to turn this into output style, tag the Claude Code guide and say, turn this into an output style.

And I may notice over time like, hey, I'm getting more technical or something like that, in which case I may want to bump up the level of the output style in some way. Now one of the nice built-in features into Claude Code is the fact that the output styles persist per project. So you can see for AgentStack over here, when I go to settings.local.json inside the .claude folder, then I have output style set to explain like I'm 5.

But if I now set the output style for this other project to the STE100 in Simplified Technical English, then I can see that the previous project's output style has not changed, but this new one over here has changed to STE100. So this is really good because you can have different output styles for different projects. Anyways, for me personally, I will be using it for different projects, different output styles. So for example, recently I have been working on a new project of building really good agent sandboxes.

So I can do /config and set that to explanatory to understand more about agent sandboxes. But for projects I'm very familiar with, I can have it be way shorter instead. Anyways, I would basically recommend playing around because I think this is something we will need to do more going forwards, as the models are getting slightly more confusing in their default outputs. Anyways, if you do like this kind of stuff, then do subscribe to the channel.

And if you do want to get more tips from me on a regular basis, then do check out my email newsletter. It will be linked down below. By signing up, you will also get access to a bunch of free videos from my agentic coding school class as well. And you're also welcome to reply to any of the issues, and I will try to get back to you as well, because I really like hearing from people replying to newsletters.