Orca: FREE AI Coding Agent (Open Source!)

Cloud Code has just created a new guide on how they actually removed over 80% of Cloud Code's system prompt for more advanced models. Now, these are based on the new Cloud 5 generation models. It's also super useful to learn this stuff if you are using any sort of AI agent. So there's multiple different rules you can take from this to really understand how to get the most out of your AI agents when you are prompting and using them.

So these are the new rules of context engineering for Cloud 5 generation models and how it all works together. Crazy thing about this is they've removed over 80% of Cloud Code's system. So there was a lot of fluff in there and they actually found when they removed that fluff, they got more out of these models. And by doing that, they actually didn't see any measurable loss on coding evaluations.

They removed like the 80-20 and actually got the same results, which is crazy in itself. So this is all about something called context engineering. It actually makes a big difference when you're using Cloud Code or any sort of AI agent. And the thing to note here is like, much of the context assembled from your system prompts, the skills that you have, Cloud MD files, the memory, other sources, this bloats up the context you actually have.

So for example, when you send a message to Cloud, that message is only a small piece of what Cloud actually reads. So behind the scenes, Cloud also gets a system prompt, your Cloud MD files, skills, memory, and all that stuff together is basically context. Now, you wanna think of it like a lunchbox for someone. So your message is sandwich, but the lunchbox also has napkins, it has a note, it has a fork, it has snacks.

And context engineering is deciding what goes in the whole lunchbox, not just a sandwich. So the tricky part is your message changes every time you use Cloud Code, but the context is reused for loads of different messages. And so it has to be general instead of specific. Now, what they actually found is that unhobbling Cloud helps a lot.

And what this essentially means is you just remove a lot of the constraints that it has. So basically less is more here. So if you have loads of rules, like for example, leave documentation as appropriate, or for example, do not add comments as a system prompt, skills, everything just kind of clashes with each other, especially if the user of Cloud Code is using something else. So let's say, for example, you're using an AI agent.

And let's say, for example, that agent has loads of rules, but then when you prompt it, sometimes your prompts actually contradict the rules that it has. The problem with that is like Cloud or any other agent, you can interpret what you're trying to do, but it has to think more carefully about overlapping and also these conflicting messages it has. And so these used to be really useful for reducing like the worst case scenarios. But actually these days, you can actually delete them, especially with the newer models.

And also there's tons of more tools inside your AI agents. So for example, if you look at Cloud Code, it used to rely on Cloud.nd. It's a source of memory, information, and guidance. But now you actually have an in-built memory and artifacts and skills.

And so Cloud has new ways of loading the context without using the Cloud.nd file. So instead of your instructions fighting each other, you can actually just use the rule less is more and over 80% of Cloud system prompt can be removed, especially for the newer models. Older models need stricter and stricter rules. Newer models have better judgment.

So many rules can actually be done. Now they've actually talked about five different rules you can see here. So give Cloud rules was the old way, but now you can give Cloud judgment. Before you'd have to give Cloud examples.

Now you can design interfaces. You can put it all up front or you can use progressive disclosure. And this is the old way. This is the new way.

So before you'd have to repeat yourself. Now you've got simple tool descriptions. So before you were giving Cloud loads of rules, but now you can actually not really worry about that too much because instead Cloud can use judgment. So like, if you look at the old system prompt of Cloud Code, it's like, okay, there were hard rules like write no comments and never write long strings.

The problem is sometimes it's just wrong. Some uses one documentation, some complicated. Code genuinely needs a larger explanation. And so the new instructions basically write code that looks like the code around it, matches new comment style, naming it habits.

So instead of a rigid rule, Cloud can just understand what code you've already got and then use that in the future. Now before as well, you'd have to give Cloud examples, but now there are design interfaces. So what they actually found with the newest models like Cloud Opus 5 and Fable 5 is that giving examples actually constrains them to a certain exploration space. So before this rule for Cloud of like show Cloud examples of how to use them, they kind of created fences where Cloud would stay inside the box.

Whereas with these new tools, scripts and files, basically Cloud has parameters that allows it to be more expressive. So before it would stay in the box like this, whereas now it has these different status options. Now before as well, it would put everything up front. Now it uses something called progressive disclosure, basically loading the right context at the right times.

Now you might be wondering, okay, how can you apply this to your own usage of Cloud, for example? Well, you can apply this to, for example, your own Cloud MD or your SQL MD files. So people before they would have like a central of like every node into because Cloud wouldn't find it otherwise. But now what you can actually have is a tree of files that can be loaded at the right time.

Because like most of the time, the information you have before was just kind of sitting there taking up space, right? Now you can have like separate skills for each workflow and Cloud would just open them when the moment calls for it, right? So you don't need like everything inside Cloud MD. There's one giant book containing everything.

You can have like a tree of skill files. And then each time, for example, it does SEO, it uses the SEO skill. Each time it uses front-end, it's gonna have a skill for front-end design. Additionally, before you'd have to repeat yourself, especially like repeated instructions.

This is really just because like, if you have a long conversation with Cloud, sometimes it would get lost on the context and not find its way back. And it would get completely off task, if that makes sense. But now you can actually just put instructions on how to use tools in the tool descriptions rather than the actual system. So what this means basically is like older models sometimes should've got instructions from the beginning of a long context.

So people have to repeat stuff here. Like if you ever have to keep like copying and pasting the system prompt, you'd have to do it again and again and again, just so it doesn't go off task. But now you don't need to do that. So you can put the instructions for a tool in the tool description and delete the copy from the system prompt.

Something else that people used to use was memory in Cloud MD files. And so what they actually used to recommend was saving things to Cloud's memory by using the hashtag hotkey to write to their Cloud MD file automatically. But now you actually just have saved automatic memories and you can also plug in Obsidian. So either way, you don't need to say like, save this, or for example, add this to your Cloud MD file.

It will remember the simple pieces as it goes along with auto memory. So something that people are using, this is something we still use inside Obsidian, is simple specs. You would have like a markdown file and those markdown files would help Cloud have a reference point when it needs to store something, particularly if it needs to store something in like code base for later. But now actually what you can actually use because they have the new artifacts feature, you can reference HTML artifacts.

So you can give Cloud references in the form of code. You don't need to use a markdown file. So that could be like a HTML artifact you use instead of a plain markdown plan. You can use code as a spec.

So a detailed test suite is a spec and so is a function in another code base and also rubrics. So they basically teach Cloud your taste, which is really important actually because like Cloud can basically code anything. But if you actually look at some of the designs that it creates, often they're quite bad and that's because it doesn't have taste. So rubrics allow you and allow Cloud to try and verify the taste in a particular field and it can use verify agents with those.

So if you were going to actually start using this, what you got the system prompt, and this is only for people building their own agent harness. So for the system prompt, that's only something you need to edit if you're creating your own harness, but the Cloud MD file is pretty important. So if you're using Cloud code particularly, you wanna keep it lightweight and just briefly explain what the repo is for, but just explain the stuff that might surprise Cloud. For example, all our types live in one giant file and nowhere else.

So you wouldn't write anything obvious inside the Cloud MD and you wouldn't add anything too obvious, you keep it super lightweight. Instead, what you wanna use is progressive disclosure. So you could, for example, like make a verification skill and just point it and point to it from your Cloud. Now also the skills here, if you're not using skills, these are super useful.

So if you run a workflow, like for example, design a page, you can say, okay, save that as a skill in the future. And these are like lightweight guides to let Cloud find information when it needs it. Now, again, you don't want to have too many rules inside there. That's actually something that I've probably made the mistake of and need to improve in the future.

We've got like loads of rules inside some of our skill files. So if you are having bought new skills inside these workflows for long skills, made them more lightweight, that's actually something I plan to use and just reduce it below so it's not too constrained. And then also for references, this again, leans into progressive disclosure, which is instead of having everything inside the Cloud MD, you add context later. So you could use app mentioned files to include them as references and then Cloud can use those later.

So you can go into Cloud and then tag in the file. And that could be like a spec file, a mockup, whatever you want. And again, it doesn't need to be like a markdown file. It could be a HTML markup.

It could be, for example, a HTML. And actually code is one of the best things you can use for references instead because it's super precise and Cloud is really fluent in code. So for example, a HTML mockup of a design will usually be a written description or a screenshot of that site. And then finally, simplify.

You can reduce a lot of the bloat inside your skills. That's something I plan to do. System prompts, Cloud MD files, you can simplify all of them a lot more. And that just reduces the constraints and gets more out of the Cloud 5 models you're using.

So thanks so much for watching. That is the new guide from Cloud and what they've changed inside Cloud Code recently to just help you get the most out of Opus 5 and also Cloud Fable 5 and any other new models that come into the 5 category. If you want to get more training on Cloud Code, how to use it, et cetera, you can go inside the AI Profitable Room. Link in the comments description or go to theaiprofitableroom.com.

This is my community that helps you save time, grow and learn with AI automation. And you can post questions inside the community. It's a great group of people where you can ask questions, get help and support. I personally answer the questions inside there and so does everyone else.

And there's always people online 24 seven to help you in if you need to. Inside the classroom, you get access to all of our best training. So if you're a complete beginner, you can go from beginner to expert with a six week course. If you want our new daily updates on the new advanced stuff, you can get that right here.

And if you want our agent OS system that allows you to orchestrate all of your AI agents, Cord, Hermes and everything else, you can grab that. And then also we have tons of trainings as you can see. You can also jump on four weekly coaching calls to get help and support real time, share your screen, ask questions. And then inside the map, you can meet other people using Cord and other AI agents locally near you.

And it's an easy way to connect with other people. Plus you can direct message me anytime you need. And we actually have over 204 pages of testimonials and wins from people learning and growing with this stuff. It's just an awesome community, great people, lots of people growing and learning and scaling with AI automation.

So feel free to get it. Link in the comments description or just go to theaiprofitblog.com. Thanks for watching.