Anthropic Just Revealed 7 New Rules for Prompting Claude 5 Models

0:00 I recently looked at Anthropic's new guides and keynotes on prompting and I learned that almost everyone is prompting Claude 5 models wrong because Claude Opus 5, Fable 5, and similar models work fundamentally different than previous models, we need to change the way we prompt in 2026. So, in this video, I'll show you seven rules for prompting Claude 5 models the Anthropic team actually uses to get the most out of these models and to avoid burning tokens. Now, if you're new here, I'm Ben. I'm a three-time founder and I now run an AI agency and one of the largest for professionals and business owners where I teach how to use AI to actually automate work. Starting with the first and probably most important one, which is giving Claude 5 models the entire job instead of prompting it step by step.

0:45 Now, both Anthropic guides on prompting best practices for Claude Opus 5 and Fable 5 mention specifically that these models tend to perform best when given the complete task specification up front and are left to run including for complex tasks. Now, the reason for this is simple. These models have been trained specifically on executing end-to-end tasks. Boris Cherny also recently mentioned this in his keynote talk at Y Combinator including a simple framework to think through when prompting these Claude 5 models.

1:15 >> I think a a really common mistake that I see is people are using Claude code, they're using Claude and they they just give it like way overly specific instructions. They're like, I want you to do this but I want you to do it in this way, this way, this way. You must do like one then two then three then four. And for modern models, that's actually really not the way to do it. You want to go a little bit higher level. You want to describe the task, you want to describe the guardrails, you want to describe like the exit criteria, and then just go with the model cook and come back in a little bit.

1:40 And I think it'll it'll surprise you. Like and again, like this is just not something that would have worked six months ago but it does work today. >> So, essentially, we want to describe the entire job that needs to be done, the why it needs to do this job, the guardrails for doing it, and of course, what done looks like. So, an easy way to help you improve your prompts for these types of models is always to just go through this prompting sort of framework when putting in a prompt. I'm a big fan of these frameworks because it just makes it easy for you to go through these best practices every time you give these models a prompt. For example, I use this prompting framework together with Claude Fable 5 to help me prepare for this YouTube video. So, I gave it a broad job or goal, I gave it the why, the guardrails, and what done looks like. I then let it run for 10-15 minutes and it came back with a pre pre-outline for my YouTube video, which was honestly extremely helpful. Now, of course, sometimes you do have smaller tasks where you might not have to be as detailed, but then using these types of models is often just overkill or you should at least use them with low effort. Now, I've put this prompting framework in the free resources in the first link in the description below together with a couple of other free resources from this video. But because, of course, we're not always 100% clear on the exact specification of an entire task beforehand, especially if it's a complex one, Anthropic recommends using a specific skill to get full clarity on the task beforehand, which brings me to rule two, which is using the interview me skill before actually sending the model off towards an end-to-end task.

3:07 So, basically, of course, when you're prompting a model for an end-to-end task or complex task, it becomes harder to consider or articulate all of the little guardrails, rules, and things it has to do beforehand. but in a recent article on how they the Anthropic team uses Fable 5 models, they do mention it's really important to articulate and know in detail what you want out of the model in order to get the best results. That's why Anthropic internally uses an interview me skill before sending one of these powerful models off to do a complex task. It's basically a skill that asks you questions about the task beforehand to get the full context and then can create a brief or even the entire prompt that you can throw into a Claude 5 model.

3:49 So, it's designed to find some of the unknowns in your task and bring them above surface. For example, here I used that skill to help me plan for a personal analytics dashboard that I want to build out for myself. And as I wasn't 100% clear on what I actually wanted or wanted in there, I used the interview me skill. So it asked me a couple of questions to get more context around my task. And then at the end, it gave me the brief, which is essentially the prompt that we can throw into a Claude 5 model to actually build the dashboard.

4:17 So you can see we have the job, the why, the guardrails, and what done looks like. Now the interview me skill will also be available in the free resource link below. But in essence, putting more effort into the pre-planning instead of quick prompting and iterating on outputs is a little bit of the mindset shift that you need to make with these more intelligent models in my experience. Because you can better spend a bit more time on creating this initial prompt and planning it. I've personally also seen this just results in better outputs from these types of models and will also help you reduce token burn and save you time because you don't have to endlessly iterate.

4:52 Now another effective way of doing that pre-planning, besides the interview skill, is what also Andrej Karpathy recommends, which is basically switching on or using a voice transcription tool and just ramble on about the task for a couple of minutes. Now in my case, I use WhisperFlow, but Claude also has a built-in voice tool. Because through your voice, you often just get far more context out to Claude and it's far faster. Now I do often notice with myself, for example, that it becomes very organized and messy, which is also not good for these types of models. So I sometimes use a prompt master skill that I built myself that basically gets a ramble like that and organizes and structures it a bit better, so I can then paste it into a Claude 5 model.

5:30 Again, the prompt master skill will also be in the free resources. By the way, if you enjoy this type of content, you'll really enjoy my AI accelerator, where I just launched three new beginner to advanced Claude courses that walk you through step-by-step how to use Claude to actually automate your work as a non-technical professional or business owner. We also have unlimited one-on-one live tech help to help you with any problems you might face and a big library of skills and pro plugins across all business departments they can use or customize for yourself. We have multiple weekly Q&As with me and my team and a community with professionals and business owners actively using AI to automate their own work or build AI business. So, if you want more information, you can check it out in the link in the description below. And if you're someone that wants a bit more help, you can also check out my AI operator program which is a 30-day one-on-one coaching program to help you get from zero to advanced really fast.

6:19 So, you can also check that out in the description below. Which brings me to the third rule, prompt why it needs to do this job, not just what. The Claude Fable 5 guide mentioned specifically that it tends to perform better when it understands the intent behind the request, especially on complex task. And the reason of course again is because on complex tasks, the model will always run into some decisions that it has to make that you haven't pre-specified. Even when you're you've pre-planned this with for example an interview skill. And when it understands the bigger picture and why it's doing this job, it can make those small decisions through better context. They even give a specific template you can apply in your prompts which again I'll add in the free resources, which is basically I'm working on the larger task for the specific person who this is for. They need what the output enables and with that in mind, you put in the request. So again, it's just a little bit of a habit you want to get into when prompting these models. For example, in that same prompt, I used that template to explain the why. So, I said, "I'm working on a video for my YouTube channel on how to prompt Claude models and how it's changed. The video's for non-technical professionals and business owners that are using Claude to automate their work.

7:25 They need practical tips, examples, and frameworks on how to improve their prompting, not just theory." And this all might feel a little bit stupid or doing too much when you first do this, but it can really improve your outputs and that's why I'm such a big fan of having these sort of frameworks, these prompting frameworks because it forces you to include these things that might not come very natural. And if you think about it, the same principle really applies to giving instructions to humans. The easy thing to do is delegating work by telling someone what to do, but the most efficient way to delegate work is also telling why they're doing it and why it's important.

7:57 So, rule four, which might be one of the most important ones, is defining what done looks like. Claude 5 models are long-running, and that means they generally don't do too little, they do too much. And that's why defining the exit criteria becomes so important, because you might have noticed it just keeps going on too long and spend far too many tokens if you don't. It's also something Boris specifically mentioned in that Y Combinator talk. So, there are really two aspects to this. First, defining what done looks like and defining the output style, which is again in that same prompting framework and is something you really always want to try to include inside of this prompt.

8:31 For example, in my case, I said a pre-outline for this video. That means eight to 15 practical tips on how to prompt Claude models together with a specific example for each and the source you found this tip for. Again, the source you found for this tip should be backed up by Anthropic. And I added, "In my second brain, you can find an example output style for a video outline." So, the model can actually get an idea of what a video outline looks like.

8:55 Here, of course, you can also add in a prompt of what kind of tone of voice it needs to use because a common problem, of course, with Opus 5 is that it can be overly verbose. So, here you can just say something like, "Give the output in plain English that's easy to understand." Now, some people have been saying even examples are not that necessary anymore, but in my experience, if you have it, it does help, especially for getting a specific output style. So, rule five is swap hard rules for reasons. Now, many people are trained in using phrases like "never do" or "avoid this", etc. But based on Anthropic's new research and article on context engineering and a keynote talk that they had on prompting, they've basically found that these new more intelligent models seem to respond far better to instructions combined with a reason behind them. So, defining guardrails is still important to do, but avoiding telling it what not to do and always adding the reason for the instruction seems to be far more effective for these more intelligent models. So, when I'm defining the guardrails for this prompt, for example, not saying never give a point that is not backed up by Anthropic. Instead, I'm saying make sure that the points that are mentioned are backed up by Anthropic's own team. So, we actually have proof for the claims we're making and the reason behind.

10:06 Again, it might sound like a small thing, but it's more just getting into the habit of changing the way you give these guardrails a rule. Which brings me to rule six, avoid double-checking prompts in Claude 5 models. Now, this same principle, of course, also applies to your Claude.md and your skill files, which is why Anthropic even launched a rule rewriter skill that helps you update your Claude.md or your skill files to swap these hard rules according to the best practices for these newer models. For example, here I applied it to my LinkedIn writer skill and it changed some of these hard rules here to instructions with an actual reason behind them. The rule rewriter skill I'll also add in the free resources link. Now, telling a model to verify its output or using a sub agent to double-check before giving you the final answer has been sort of a popular prompting technique for trying to reduce mistakes and hallucinations. But Claude models like Opus 5 and Fable 5 have been specifically trained to double-check and verify and fix mistakes autonomously, meaning that specifically instructing this just adds another unnecessary verification step to the process, which adds cost without really improving results. Now, besides that, there are few other popular prompting techniques that the Anthropic team recommends avoiding with these newer models, including the think step-by-step, explain your reasoning, or using aggressive emphasis like this with capital letters because it can cause over triggering. Now, if you want to look at all these rules in detail, the list will also be in the free resources. And the last one applies to a problem many people have been facing specifically with Claude Opus 5, where the model starts using a lot of jargon, becomes overly verbose, and it sort of becomes unreadable, which is basically a fix Claude's voice once rule. Anthropic essentially recommends using a simple prompt inside of, for example, your Claude.md or your Claude desktop global instructions or your system prompt to fix any voice issues you might have with Opus 5. They give an example in the guide again, but of course you can customize this and specify it to the specific output style you're looking for. In the Claude desktop, for example, you can just go in general and in instructions for Claude, give the specific output style you're looking for. In my case, keep responses focused, brief, and concise, avoid jargon, and being overly verbose. But you can also, of course, add this in your Claude.md or even in your system prompt if you may be using projects.

12:27 So, essentially, all these rules really come down to changing a little bit of the mindset and the way you think about prompting some of these newer models. It becomes more and more essentially like instructing a human. You don't just want to tell a human what to do, you want to tell them why they'll need to do it, and they'll often do their work better. And I think Anthropic put it very nicely in one of their articles, "Think of Claude as a brilliant but new employee who lacks context on your norms and workflows, and the golden rule is show your prompt to a colleague with minimal context on the task and ask them to follow it. If they be confused, Claude will be, too." So, essentially, we have to start prompting these models more and more like humans. Again, if you want to dive deeper into AI, definitely check out my AI accelerator in the link in the description below. And if you want to learn more about how to use Claude effectively, you can check out the video here above.