I Analyzed How Anthropic ACTUALLY Prompts Fable 5.1

0:00 Today I've got four tricks from Enthropic themselves on how to get more out of Fable 5.1, how to use it more efficiently, and how to stretch that weekly usage limit as long as it can go.

0:08 So, let's not waste any time and just get straight into today's video. Okay, so before we hop into these things, I wanted to say that I really am liking Fable 5.1 so far. I will be honest, as

0:16 of lately, I've been using Codeex a lot more than Claude Code, but Fable 5.1, it's feeling really nice. It's feeling pretty quick and it's feeling pretty efficient if you learn to use it right

0:26 because I have seen tons of people complaining that they are just killing their session limits. So, let's get into it today. Now, everything that I'm talking about today, I found straight

0:32 from the Claude platform docs on how to prompt Claude Fable 5.1. There's a ton of gold in here and it talks about every individual model because these are different models. They have different

0:42 characteristics and they perceive information a little bit differently. So, it's always worth checking out. There's a ton of info in here. Like, this is a long long doc with lots of

0:50 scenarios. So, I dug through here and I just pulled out what I think you guys actually need to know to start using it differently right now today. So anyways, the first one, tell it what done looks

0:59 like. This means you're not giving it tasks, you're giving it the finish line, and you're letting Fable 5.1 figure out the tasks that need to come together to create the end result. I've heard Boris

1:08 Churnney, the creator of Claude Code, say so many times to give these models an ambitious goal, and get out of their way. So tell Fable the outcome, tell it why it matters, tell it what done means,

1:18 and tell it if there are any real constraints. So instead of saying something like out. >> So instead of saying something this

1:30 wordy, could you just shorten this down into one overall goal without defining all the individual tasks? Boom. Now we have create an appealing landing page offering voice agent

1:41 solutions tailored to our target audience. And then it will find out what it needs to do. Now, for all of these tips, what I'm going to do is show you guys real quotes and real evidence from

1:48 this documentation where I pulled these tricks from. But what I did want to show you real quick is that some of these will be coming from the Claude Fable 5 docs. But I wanted to show you guys this

1:57 right here. Your existing CloudFable 5 prompts should perform well on Claude Fable 5.1 without changes, but a handful of behavioral differences are worth knowing about. So this document

2:06 prompting Claude Fable 5.1. A lot of this is just like in addition to the way that you prompt Cloud Fable that you can find in this doc. So I just needed to make that quick disclaimer as you can

2:14 see. But anyways, let's get into some of this evidence here. So this first piece, Claude Fable 5 tends to perform better when it understands the intent behind a request. Context lets it connect the

2:24 task to relevant information rather than inferring intent on its own. So provide context about why you're asking, especially for longunning agents drawing on multiple work streams, which is why

2:33 building out your own AIOS that has information about context, your your goals, your background, that is so helpful in driving powerful models like this. This next one here, refactor

2:42 existing prompts and skills. skills developed for prior models are often too prescriptive for Claude Fable 5 and Fable 5.1. Meaning, if you have some skills that are literally like, do this,

2:52 this, this, this, this, like so, so specifically, that is kind of the way we were taught to build skills, but that might actually be getting in Fable 5.1's way where it's making it less efficient

3:00 and it's kind of just like putting guard rails on it. And the whole idea of like rerunning your skills for different models is really real. It's something I've said before and I wanted to show

3:08 you guys this this tweet that I saw today from Peter Yang. If you're trying out Fable 5.1, I highly recommend running/claw- API prompt- audit. Run it on your skills. It finds a bunch of

3:17 redundancies and rules to remove for the latest models. So, go ahead and try that. And then this last piece here was about finishing the whole task. Cloud Fable 5.1 can execute very long tasks

3:26 without much guidance on methodology, especially when the goal is clear. But that's the key there. The goal has to be very clear. So, that is why tip number one was to tell it what done looks like.

3:38 Do it in a way where you're not giving it task after task after task. All right. Number two, we have to match the effort to the task. How many of us just hop on a claude and we don't actually

3:46 drag the effort slider around and we just stay at high? Because as you guys know down here, when you click on the model, you obviously choose that there. And when you click on the effort level,

3:54 you can change this from high to medium to low to extra to max or to ultra code. And by default, Claude Fable 5.1 just sits on high. But take a look at these benchmarks, which is really interesting.

4:08 Over here, you can see Fable 5.1. Let's just look at the version with no tools. And you can see Fable 5 with no tools. And the point I'm trying to make right now isn't that 5.1 is much more

4:16 efficient or better than five. The point I'm trying to make here is that it really differs. Fable 5.1 on low is so much different than Fable 5.1 on max. Fable 5.1 on low is about comparable to

4:25 Fable 5 on medium or high, but it's also cheaper. I've also heard people compare Fable 5.1 on low to things like Opus or sometimes a really, really good sonnet. Now, the point I'm trying to make here

4:34 is the majority of work that you're probably doing with Fable 5.1, Fable 5.1 on high is overkill. I see this meme on Twitter where it's like this guy is lighting a cigarette with a blowtorrch

4:44 and it's like using Fable 5.1 to reorganize my downloads folder. Like the normal model nowadays, Sonnet, you know, GBT 5.5, like these older models are still really, really good. Like remember

4:54 when Sonnet 3.7 was like the best thing out there? It's still a really good model. It's just that we have things that are so much better now that that doesn't always have to be the default.

5:02 So, the point I'm trying to make here is do not automatically run everything at maximum effort or even just high. Start with high, then test whether medium can do that same task just as well. And if

5:12 it can, try it on low. And you only need X high and max when you're doing like massive massive work or something that needs like deep deep thought. And obviously, that kind of stuff is going

5:20 to cost you a little bit more as well. Okay, so let's take a look at some evidence here. The first one is to consider all effort levels. Pretty much just what I said. Start with the default

5:27 at high, then test on other levels like low, medium, X high, and max against your own EV valves is what I think is really clear because some people are like, "Oh yeah, I drive all the time

5:36 with X high and it's it's perfect for me. Anything lower and it's not as good." Well, maybe they're doing like deep software engineering and, you know, merging 15 PRs every second. That's

5:46 probably something that does need X high. But maybe you're creating documents and creating spreadsheets that probably doesn't need X high or high. So, it has to be on your own evals and

5:53 your own use cases. The next one we have here is search triggering at low effort. So just think about this. At low effort, CloudFable 5.1 is less likely than CloudFable 5 to call a search or

6:04 retrieval tool and more likely to answer from memory. So there are lots of decent use cases for switching CloudFable 5.1 to low when you still want that Fable 5.1 feel, but maybe you don't need to be

6:14 doing all of this tool calling as much and you're maybe ideulating, brainstorming, searching for things. And then look at this about changing effort mid-con conversation. You can run later

6:23 turns of a conversation at a different effort level in two ways. Now on Claude Fable 5.1, you can use a per message effort change which keeps the prompt cache. But on other models, you have to

6:33 set a new top level on the new request. So Fable 5.1 lets you switch effort in between questions, in between texts, which is pretty cool, in between turns. All right, moving on to number three

6:43 here, which is my favorite one of all time, not just for Fable 5.1, but just in general working with AI, is to make it prove its work. So think about how would you check this output if a human

6:53 came to you and gave it to you? What would you do? Would you just look at it like a website? Would you watch the video? Would you try to like look up the sources and fact check? Would you test

7:01 the UI by clicking? Whatever you would do here, apply your own taste, apply your own judgment. Let the AI models do it first. The whole idea is you don't want to get handed a rough draft. You

7:11 want to get handed a version five or a version 10 that's already been approved and reviewed by lots of little sub agents that do verification and then it's been iterated on like five or six

7:20 times or maybe even 10 times because every single time they review it, they find a mistake, they fix it, they review it again, they find a mistake, they fix it. Now the tough part is not every

7:28 single piece of output can have an objective check. Like if you were literally trying to optimize the performance of something so that X equals 10, that's something that

7:36 objectively can be proven. A lot of these sorts of outputs need to be verified from more of a subjective level, which is really you creating your own sort of like LLM as a judge, which

7:45 means that you do have to put your, you know, input in there as far as this is what good looks like. So, you define the goal, but you also need to very firmly define what does good look like and how

7:54 do you test if X is good. So, it's not going to get you 100% of the way there, but you'd rather get 97% of the way there than getting like 70% of the way there. You know what I mean? And here's

8:03 some evidence from the docs. You can see this first one is make self-verification explicit in long run prompts. So literally telling it here verify your work with sub agents against the

8:12 specification. Down here another one you can see in the prompt it says verify your work however you like. And then down here we have um before reporting progress audit each claim against a tool

8:21 result from the session. Only report work you can point to evidence for. If something is not yet verified, say so explicitly. And then just on the vision piece, because I know a lot of us are

8:30 building websites or slide decks or whatever it might be where you need some vision verification, Claude Fable 5.1 has better vision capabilities out of the box and on complex visual inputs

8:39 such as dense charts. It does its best work when it can iteratively analyze, crop, and visually verify what it sees as it goes. And finally, moving on here to number four, we have parallelize and

8:50 delegate. So breaking a large task into independent tasks. Just because you have one process doesn't mean one agent needs to be working on it. You can have one process where you actually have 80

9:00 agents each go in there, pick one piece of it, like an assembly line, and they hand stuff off to each other, or they all do their own piece at the same time, and then it all comes together in the

9:08 end. Not every task is available to be like parallel work, but a lot of them are. So, this will give you faster completion, better focus and specificity because each agent has one very specific

9:19 job now and they'll just they're going to do better broader coverage and context efficiency because now you can keep that main session cleaner and conversate with you while you have it

9:29 delegate out to a bunch of different agents. And this is like the most important thing that I've been able to do to get way more usage out of my Fable 5.1. Most of the time now when I'm

9:37 working with Fable 5.1, I tell it, I don't want you to build anything. I don't want you to code anything. I don't want you to research anything. I just want you to spin up sub aents, drive

9:47 strategy, interpret what they give you, and then spin up more sub agents because now we're getting Fable's intelligence to do that, and we're saving so many tokens. Trust me, just try out using

9:56 some prompts like that, and you will feel the difference immediately. But that doesn't mean that Fable isn't responsible for the outcome. Like, if it's a bad output, then it was Fable

10:03 5.1's fault, not all the sub agents faults. So, it needs to keep in mind that it still has to do things like spinning up verifier agents or maybe at the very very end on the final pass

10:11 before it gives it to you the human. Maybe that's when Fable 5.1 actually steps in there and it gets its hands dirty a little bit at the very end. Anyways, let's look at some evidence

10:18 here. So, the first one, delegation and collaboration. Cloud Fable 5 is significantly more dependable at dispatching and sustaining parallel agents and reliably manages ongoing

10:27 communication with longrunning sub agents and peer agents. Then we have batching independent tool calls. Fable 5.1 usually issues parallel tool calls as expected. When a request names

10:36 several things to fetch, it issues those calls all in parallel so they can all work at the same time and you're not just sitting there wasting time. And once again, parallel sub agents save

10:44 time and cost through cache reads and avoid bottlenecking on the slowest sub aent. If you got an assembly line and agent number three is just taking forever, then 4 5 6 7 8 9 10, all of

10:54 those other agents are just sitting there and waiting. So if you can parallelize them, parallelize them. So, I know that was a decent amount of information that we just covered and I

11:01 put all of this information with all of this evidence into just a super simple resource guide that you can access for completely free. All you have to do is go to my free school community. The link

11:09 is down in the description. You'll come into here, you'll click on classroom, you'll go to all YouTube resources and every free doc or skill or repo or guide, whatever it is, I've always given

11:19 them away for free right inside of here. So, that is going to do it for today's video. I hope you guys enjoyed or learned something new. And if you did, please give it a like. It helps me out a

11:26 ton. And as always, I appreciate you guys making it to the end of the video. I'll see you on the next one.