Hey thanks for watching! I'm Jay - spent my career in data and brand building, founded the ROBO Group to help forward-looking businesses grow with AI, and now teaching what I know through this channel and the RoboNuggets community.
This video provides strategies to combat the verbosity and technical jargon often found in high-end AI models like Opus 5. By utilizing custom \"output styles\" and \"skills,\" users can reduce token waste and improve readability, shifting the focus from raw benchmark power to a more practical, user-centric experience.
There's a widespread problem plaguing Cloud users recently, specifically on Opus 5,
because its replies tend to be confusingly long, packed with all too plausible nonsense,
and it's becoming more and more jargon-dense. The good news is that Entropic already shared
a fix for this. It's super easy, and the principle behind it makes your setup faster
and cheaper, and it lets you finally understand what the heck your agents are actually saying.
And by the end, I'll hand you a resource that solves all of this automatically,
just by sending it to Cloud. Let's dive into it.
If you've been using Opus 5 heavily the past few days, you may have noticed two specific
problems with it. The first one's the jargon problem, which is its tendency to use technical
terms or acronyms much more versus earlier models, which makes it much harder to understand and work
with. Just to give you one simple example, and this is probably not the worst of it,
if you ask a simple question to Opus 5 like this, where I said our CRM dashboard says open rate is
32%, and I said from a conversation with a peer, we discussed how open rates are basically fake
now. So I asked it, what does that mean, and should I care? And it did give us a proper
explanation here, but it seems to always default to being more comprehensive and more technical
than most people would probably want it to. And you can see its tendency to assume that you know
every single jargon in the field as well. Like for example, here it's saying every Apple mail
recipient with MPP on registers as an open, and it also tends to use a lot of technical jargon
like send time optimization, geo data, or device data derived from opens. MPP masks IP and reports
a generic device, so all of this is corrupted. So you can see what I mean, right? And like I said,
that example was probably not the worst of it, because I first became aware that this is sort
of a widespread problem when I saw this tweet by Peter Levels, who's pretty popular in the
hacking space, where he shared this snippet of an article that made the rounds around X recently.
And I can probably show the actual article just to give credit where it's due. So this was written
by Niklas Grunn, if I'm pronouncing it correctly. And I think he summarized the problem really well
here, where he's saying that reading AI output today is extra effort. You can pretty much do it,
but it is extra effort. It's verbose, it frequently contains all too plausible nonsense,
and is increasingly jargon dense. And then he says that he recently got this sentence from Claude that
even I don't know what this means, really. And then the second but related problem is the wall
of text issue. So if we go back to our example here, you probably experienced the same thing
where you ask it a simple question, and then it goes into a full blown essay and gives you a wall
of text as an answer. And for most cases, there's multiple issues with that, right? Because it wastes
your time, because you need to understand and read through this whole thing in order to get to the
point. And secondly, there is also a corresponding cost for tokens, because the longer its output to
you, the more tokens that Claude is consuming. And so if you have this same problem with Opus 5 as
well, which a lot of people seem to be using as their default, then you're in the right place.
Because today, I'll be showing you the two solutions for those two unique problems. And by
the way, if you want to learn how to build and sell AI systems that businesses actually pay for,
then that's pretty much all we do over at the Robo Nuggets community, where not only do you get
access to the Claude living masterclass, which we update every week and takes you from zero to
mastery with the latest on AI, but you also get access to our agents as a service course,
which walks you through how to actually get paid for all these AI skills that you are learning,
you also get to be part of a genuinely great community of AI builders. In fact,
you can see just some of the recent wins our members are getting from the program right here.
So if you want to start earning from AI, then check that just in the pin comment below. Now
back to the video. So let's start with the jargon problem. And the good news about this is that
entropy steam actually talked about how to fix this one issue. And that fixes courtesy of Lydia
is a member of technical staff at entropy. And essentially, it all has to do with this
configuration in cloud code that's called the output style. And the output style is simply
just that it is the style by which cloud code provides its output to you or how it talks to you.
And I'll just show you where to find it, at least in the terminal view. But I'll show how to do it
in the desktop app as well if in case you're more comfortable there. But here in the terminal,
if you have cloud code open, you can slash config, which brings up all your settings.
And if you find output style here, and if I click on enter, what you would most likely find is that
your cloud is set to the default output style, which is that verbose jargony output that Opus
five usually defaults to you have other options here like proactive explanatory and learning that
cloud code ships with by default. But you can see that I have a fifth one here that I added as a
custom called Eli five. And this Eli five is the fix that Lydia was talking about here. And in case
you're curious Eli five simply stands for explain like I'm five, it is a popular sub Reddit. And so
the reason why that acronym works for a lot of models, including Opus five is because they've
been trained with the entirety of the internet, including that subreddit. If you're curious,
this is what that subreddit is. And there's like 1000s of posts in here asking questions in the
Eli five format. And you can be sure that these models have scraped this whole thing in order to
understand how people respond to these questions. But the core principle of it is if you want things
to be simple to understand, then this Eli five is one way to do it. Now, obviously, you don't
have to adopt the exact verbiage that Lydia has here. For example, one thing that I added in my
output style personally, is this tip from Andrew that I found where he asked Claude to only report
to him in this ASD STE 100 simplified technical English. So what is that? Well, in a nutshell,
that's essentially a controlled language standard that uses a more restricted dictionary. So a list
of words that are much easier to understand. So it avoids jargon, it avoids a lot of those technical
vagueness. And at least from what I found personally, and for a lot of people, it seems
that standard seems to help a lot in terms of just making talking to your agents much easier
and much more seamless. So how do you now apply this fix and actually change your output style
so you can test it out? Well, what you can do is just copy this prompt or take a screenshot of it
and send it to your cloud code. And what you're saying here is simply to set your output style to
the one below, which is this whole snippet from Lydia, with that addition of the ASD STE 100.
And then the important part here is the direction to save it in your output styles folder, which
obviously depending on your workspace, that location of that folder may be different. But
the good news is cloud code can probably find that for you. At least for me, this is where it put
that markdown file. And within this markdown file is just this piece of text. And then the second
thing that this prompt will do is to just set this output style for Eli five in the appropriate
settings files. And you can see for me because I already have it set, that is what cloud also told
me where it created that style file for me that Eli five dot MD, and it also updated the relevant
settings. And the best way for you to check is to just open up a new session and ask what output
style it's using now just to confirm. And if you need to make any changes to it, then you can just
tell that to cloud and it will update this output style for you. Now you may be asking why it's
important to put these rules in the output style instead of something like the cloud dot MD. And
the truth of the matter is that you can actually put these rules into cloud dot MD, but it just
isn't as effective. And if you're curious, it has to do with two things. One is the strength of the
placement with regard to the output style, because if you notice what we changed earlier is the
actual settings for cloud code. So it's written into the core system prompt of your cloud code
instance. And then more importantly, at least I would think is that under the hood, what cloud
code actually does is it auto injects these sort of not just these reminders, like stick to your
output style, even midway throughout your session. So if you were to show that visually your cloud
dot MD, that does get preloaded into every session that you start, but only at the top, right? But
the great thing about putting this rule into your output style is that it becomes part of the core
system prompt number one, and number two, throughout your session, cloud code actually gets reminded
about this rule so that it follows it more consistently and religiously. And just to give
an example of that same question that we ask it, if you read through this, it's much more legible
now, at least in my view, it leads with the key answer where it's saying that that open rate is
not all real people. And then it tells the clear story here of why that is, and then ends with this
conclusion that anyone can understand. Now that fixes the jargon problem. Now let's go to the
second issue, which is this wall of text that you usually get from Opus five. And for a lot of use
cases, especially knowledge work or production code, the tricky part about this problem is that
you do sometimes need long pieces of text in order to get a lot of the detail that you need. And so
this is why generally I won't advise people to put the wall of text fix right in the output style.
And so what I would generally recommend is to just use a skill when you need that wall
of text to be shorter. And in our community, I always advise people to make their own skills
because each of our work is different. But let me give you some ideas and also share how I do
it personally. One really interesting skill that I found is this one called slash pro. And it's
probably one of the shortest, but really useful skill that you can add to your arsenal, because
it is only one line. So its sole purpose whenever you invoke it is for your agent to restate the
last message in plain human language with zero jargon. And so if we go back to this wall of text,
and you can see here, I just use that skill slash bro, when it rewrote their response is saying
here, okay, so basically that 32% never met 32 out of 100 people actually read this. And here's how
it actually works. So I'm not going to read the whole response here. But you can see just how much
more understandable and better it is versus what we started, right? Another example of a skill that
I found is this one from Matt Pocock. That's called Wait, what? So again, it's a very short
skill, it says I don't understand where you've got to hear repitch that and give a little bit of
context. And this is what I'm saying where you need to create your own skills. Because you can
see from Matt's workspace here, he probably has this context.md that he uses a lot, you may not
have that. So what you can do is to take a screenshot of this or copy this skill and just
ask Claude to personalize it for your own setup. Now what I tend to do personally in conjunction
with slash bro, since you can see sometimes Opus five still defaults to a wall of text, if it
doesn't know how long you want the response to be is I use this skill of mine called slash quick
where I just declare a number in this case three, and it gives me just three points in sequence of
what it thinks is the most important takeaway from that whole essay that it wrote. And just to show
you what that skill looks like, I think it's this one. Yeah, so you can see this is the slash quick
skill. And you can see just in its current state, it just has rules on when to use it when I don't
declare a number when I declare a number. And apparently there's now how many two modes here
that probably naturally arose since I use this skill a lot. And I also asked Claude to tweak it
depending on our conversation. But that is one idea by which you can cut that wall of text into
more bite sized chunks. But like I mentioned, it's probably better for you to just get these ideas
for skills and create them yourself as well depending on your type of work. Now what I
talked about just now are pretty simple fixes. But if you take a step back, there's actually a
larger takeaway to this whole thing. Because interestingly, and if you remember, Opus five
is actually the best in terms of a lot of benchmarks that we see online, including this
one from artificial analysis dot AI where it is leading at the moment. And so the first key takeaway
here is that even though benchmarks show the raw intelligence of these models, user experience can
sometimes differ. So it's sort of like if you bought a computer with the highest specs available,
but then you actually find that you prefer your older computer just because of maybe some unique
features that that old model have. So it's quite similar to AI models also is what I'm finding.
And so what you should realize here is that let's say a year or two down the line where you have
Opus seven or Opus 10 already, if those future more powerful models have the same issue as well,
the good news about it is that these AI models are very malleable and you can just tweak your
agentic operating system or essentially the way that you talk to these agents through these same
solutions that we talked about. And that will just provide a much better experience for you
when you work with agentic AI. And so just remember that you can always direct these models with your
preferred output style, as well as these skills, which are essentially shortcuts. And just to make
it easy, everything that I talked about in this lesson in this video, I just put it all in this
eight page PDF, which you can probably just send to your cloud code and just pick out the prompts
and the styles and the skills that you would need and just tweak them to whatever your workspace
needs. So you can just find this whole thing down in the description. I hope that was helpful. And
as always appreciate you for making it until the end because that helps me a lot and I'll see you
next time. Thanks