Anthropic 刚把价值 $300 的 Prompt 工程课程直接降到了 24 分钟，而且完全免费。这套由官方开发者亲自授课的实操指南没有注册门槛，前 8 分钟讲的内容就足以击穿市面上大多数付费课。 别再去买那些昂贵的废话了，直接看官方怎么定义 Claude 的正确用法。 https://t.co/XUR8b2RjOG

Hey, everyone. Thank you for joining us today for prompting one of um my naamehana. I'm part part, the a appliday istem here in anthropic and with me as christian, also part of the applied i tem. Um we're going going do do today is take through through a little bit of prompting best practices, and we're going to use a real world scenario and build up a prompt together. Um so a little bit about what prompt engineering is um prompt engineering. You're all probably a little bit familiar with this. This is the way that we communicate with a language model and try to get it to do what we want. So this is the practice of writing clear instructions for the model, giving the model the context that it needs to complete the task and thinking through how we want to arrange that information in order to get the best result. Um so there's a lot of detail here, a lot of different ways. You you want to think about building out of prompt. Um and as always, the best way to learn this, it's just to practice doing it. Um so today, we're going to go through a hands on sceno. Um we're going going to use an example inspired by a real customer that we worked with. So we're modified what the actual customer asked us to do. But this is a really interesting case of trying to analyze some images and get factual information out of the images and have clad maica judgment about what content it finds there. And i actually do not speak the language that this content is in, but likely christian and cloud both due. So i'm made a positive to christian to talk about the scenario in the content. So for this example, that we have here, it's intended. So to set the stage. Imagine you working for a swediinsurance company and you deal with a car insurance claims on a daily哪哪儿？ Um and a ppose of this is that your two pieces of information we're going to these and details, well, but visually, you can see on the left hand side, we have at a car accident report form and just detailing out what conspired before the action accident actually take place. And i find ally meit. It's a human joone and a sketch of how the accident took place as well. So these two piece of information is what we're going to try to pass on the clb. And to begin with that, we could just take these two and throw them in to concile and just see what what happens. So if we transition over your council as well, we can actually do this in a real matter. And this case here, you can see we have our shiny, beautiful and thropeid concile. We're using the new called four soil models. Well, in this case, selling temperatzero uh and having AA huge maxsopking budgejeiles is helping else make sure that there's no limitations to what i can do. In this case, you can see a very simple problem to setting the stage or what colds supposed to do in this case, mentioning that this is intend to review it and acts and report forum, and eventually also determined um what happened accident and who said fault? So you can see here with is very simple problem. If i just run this, let me got to preay you.呃，we can see here that cloud things that this is in relation to skiing accidents that happened on the street, cold, sharp on authanerary, commmonstrating, sweden. Um and in many ways, you consiort of understand this in is a mistake in a sense that in our problem, we actually haven't done anything to set the stage and is actually taking place here. So it's sort of first guess, is not too bad, but we still notice lot lot intuition that we can bait into thought. So we switch back to slides that can see here that um we many know some best pracering is very intractive and poiracical science. Um this case here we could almost have a test cture. And the cw is supposed to make sure it understands is in a car or rehaiclar environment. Nothing to do skiing. Um and in that way, you intrtito be build upon of your prompt to make sure it's actually tackling the problem you're intending to solve um and to do so. We go through some best practices of how you we we and don't to rop a great cture is and interntly and how we recommend others to do so well. So we're going to talk about some best practices for developing a great prompt. So if 're, we want to talk a little bit about what a great prompt structure looks like. So you might be familiar with kind of interacting with a chap bot with laud going back back forth, having a more kind of conversational tional stle to interacwhen. When we're working with a task like this, we're probably using the api and we kind of want to send one single message to clad and how it nail the task, the first time around without you need to uh kind of move back and foring. And so the kind of structure that we recommend and setting the task descption tion fronfronend telling claud, what are you mto to do? What's your role? What task are you trying to accomplish today then we provide content. So in this case, it's the images that christian is showthe the form and draraying of instrucition and how they occurred. That's our dynamic content. This, that also be something you're 're trieving from from another system, depending out what your use case. We're going to give some detailed instructions to claude. So almost like a step by step list of how we want cloud to go through the task and how we want it to um tackle the claud. We may give some examples to laud. Here's an example of at some piece of content you might receive. Here's how you should respond and how we that content对。 And at the end, we usually recommend repeating anything that's really important for cllad to understand about this task, kind of reviewing the information with cloud, emphasizing things that are extra critical. And then telling cloude, ok, go ahead and do your work. So here's another view, this says a little bit more detail a little bit more of a breakdown, and we're going to walk through each of these ten points individually and show you how we build this up in the console. So the first couple things christian 's going to talk about the task context. And the tone context perfect.所以i会begin with the task context as he realized when i went through lio demo there. Um we didn't have much ellaborating work. What what's scenario clothes actually working within? And because that you can also tell their call is n't necessarily, he's a guess, a lot more or much actually want from it. In our case, if we want to break that down and make sure we can give more clear cut destructions um and also make sure we understand what's the teask. We're asthink logic.对um second as well. We also make sure we had lived a tone and title. Um key thing here is one call to say factual and to say confident. So if uccl can understand what is looking at, we don't want it to guess. And just sort of misleus, we want to make sure of any assessment. And in our case, we want to make sure that we can understand who's a fault here, well, make sure that assessment is as clear and as confident as possible, if not worth of using track, what were doing. So we transition back to the the council. Um we can jump to AV two that we have here. So i just navigate me too. You can see here. Um i also just illustrate the data because we didn't really do that last time around, just really high like what we're looking at. So we're seeing here. This is that car accent report form. And it's just seventeen and different checkbox es going into what actually happened. You can see there's a vehicle a and vehicb both on the left that right inside. And the main purpose of this is that we want to make sure that clothto understand this manually generated data to assess for actually going out了。 That is craworate to by if i now get back here to the sketch that we can hide it here as well. In this case, the form is just a different data point for the same scenario. Um in this case, here, we want to baak in more than information into our version, too. Uh, i'm by doing so, i'm actually elaborating a lot more and what's going on. So can see here, specifying the this AI systems was to help a humans claim claim the justice as reviewing car action report forms in swedish as well. You can see here. There are also allowering that as a human drawn sketch of incidents, and that should now maan assessment. If it's not actually fully confident. I that's really key because if we run this, you ll see that and you can see the same settings as well close for a new shining model, zero temperatures as well. Who run this, we can see here what, and she happens. In this case, um close able to pick up that uh now relay into car accidents, not skiing accidents, which is great. You can see is able to pick up the vehicle. A was martin on checkbox x one. And in the ecube was on twelve. Um and if we scroll down though, we can still tell that there's some information missing for clothes, make a fully confident determination of who's have fault here. And this is great. This is pretained to your task ter or set, make sure you don't make anything any claims aren't um uh, factual and make sure you y'need stort have set things when you're in your conference. But there's a lot of information were still missing here regarding the form what the form actually entails and a lot of informations. We want to want to back into this. Hello, i'm application as well in the best way of doing so is actually adding it to the system problem, which chineble is elaborate on. So back in the slides, we have the next item we're going to add the prompt. And this is um background detail, data documents and images. And here, as christian was saying, we actually know a lot about this form. The form is going to be the same. Every single time, the form will never change. And so this is a really great type of information to provide a clad to talk laud. Here's the structure of the form. We'll be looking at. We know that will not ever alter between different queries. The way the form is filled out will change, but the form itself is not great page. And so this is a great type of information to put into the system prompt also a great thing to use prompt cashing for you're considering using prompcaching. This 'll always be the same. And what this will help lad do is spend less time, trying to figure out what the form is the first time. It sees the form each time. And it's going to a do a better job of reading the form, because it already knows what to expect there. So a everything i want to touch on here is how we like to organize information in prompts. So cloud really loves structure love's organization. That's why we recommend following kind of a standard ds cture ture your your mpms. And there's a couple other tools you can use to help quad understand the information better. I also just want to mention all of this is in our dogs with a lot of really great examples. So definitely take pictures. But if you forget to take a picture, don't worry, all of this content is online with lots of examples, and definitely encourage you guys to check it out there, too.嗯，anyway, the so some things you can use the liitters like xaml tags, also markdown is pretty useful to cloud, but xaml tags are nice because you can actually specify what's inside those tags, so we can tell cloud. Here's here's user preferences. Now you're going to read some content, and these exml tags are letting, you know that everything wrapped in those tags is related to the users preferences, and it helps cald refer back to that information, maybe at later points in the prompt. So i want to show in the back in the console, how we actually do this in this case, and christians gonnpull up our version three. So we're keeping everything about the other part of the user prompt the same. And we've decided in this case, to put this information in this system proms, you can try this different ways. We're doing it in the system promphere, and we're going to tell cloud everything it needs to know about this form. So this is a swedish car accident form the formmobian swedish. It'll have this title. You ll have two columns. The columns represent different vehicles will tell claud about each of the seventeen rose. And what they mean you might have noticed what we ran up before. Cloud was reading individually each of the lines to understand what they are. We can provide all of that information up front, and we're also going to give cloud a little bit of information about how this form should be filled out. This is also really useful for cloud. We can tell it things like, you know, humans are killing this scar about basically, so it's not going to be perfect. People might put a circle, they might scrible. They might not put an acts in the box. There could be many types of markings that you need to look for when you're reading this form. We can also give a lot a little bit information about how to interpret this or what the purpose or meaning of this form is. And all of this is context, that is hopefully, really to help cloud do a better job analyzing the form. So if we run it, everything else is still the same. So we've kept the same user prompt down here. Odior scis is backwers from mine um h. We have the same user prompt here still asking cllad to do the same task, same context, and we'll see here that it's spinless less time kind kind of nraalto us. So little little it less about what the form is because it already knows what that is. And it's not concerned with kind of bringing us that information back. It's going to give us a list st, what what it found to be checked. So the sketch shows, and here lad is now becoming much more confident with this additional context. So we gave the claude lalad now feels as apapproate this to say, vehicle be was that fault in this case? Based on this drawing and based on this sketch. So already we're seeing some improvement. And the way clad is analyzing these. I think we could probably all agree if we looked at the drawing and at the list that vehicle be as fault. Um so we like to see that uh. So we're going to go back to this slides and talk about a couple of other items that we're not 're seeing in this prompt, but can be really helpful to building up building up your problems and making it work better比赛。 I think one thing that we really highlight is examples. I think examples or few shot is a mechanism. The really is powerful in steering cloud. So you imagine this um in quite a nontribual way as well. So imagine you have scenaris situations, even in this case, concrete accidents have happened that are um tricky for calling to get right? But do you with your human tuition and your human legal data is able to actually get your right conclusion that you can bake that information into the system from yourself by having here cut examples of a the data at lesppoast. Look out, you can have masual examples. You can use basically foreign, colder, AH um an image and have us part of the day if you're passing along into examples. And then on top of that, you can have the sort of depiction or description rather of how to break that down on the senate. This is something really highlight and and emphasize in how you can stripush the limits of your elean. Appappication is by bacon in these examples, inter system from. And this again, is sort of empmpirical science of prom engineering that you sort of ve always want to push the limits of your application and get the feedback loop in where is going wrong and try to add up the susisting problems that next time. One example is sort of linux that pacspace. It's able to actually referencit in its example, said, you can see here as well. This is just civil example, how we do this again, really emphasizing the sort of exmell structure that we we we enjoy it. It's gives a lot structure the call as what has been finine tuned on as well. And it works perfectly well for this example. In our case, we're not doing this just because it's is simple demo. We can realistically, imagine if you were building this for an insurance company, you would have tens, maybe and hundreds of examples are quite difficult, maybe in the gry that you'd like to make sure that call actually has some basis in to make the verdict next time. Another topic. We really want a highlight light, which we're not doing this. Stemo is conversation history. It's in the same being in as examples. Um we use this to make sure that the enough context rich information is that close the poposal, when when, when lworking on on on your behalf. Um in our case, now this isn't really user facing at a leunnification. It's more something happening. The background you could imagine for the insurance company. They have this automated a system, some datas generated out of this, and you might have a human, the loop at twourse end. If you are to build something much more user facing, where you'd have long, long commization history that would be have relevant to bring in. This is a perfect place in the system. Prompt include within um our rich is the context that cloud works within um in our case, we haven't done so. But what we do is and the next step is try to make sure we give a concrete reminder of the basket hand. So now we're going to build out the final part of this prompt for cloud. And that's coming back to the reminder of what the immediate task is and giving cloud or reminder about any important guidelines that we want it to follow some reasons that we may do. This are a preventing illusiinations. So we want cloud to not invent details that it's not finding in this proportrit or not finding in the data. If clad can't tell which form is checked, we don't want clad to take its best guess or invent the idea that a box might be checked when it's not. If this ketch is unintelligible, the person did a really bad job drawing ing this drawing, and even a human would not be able to figure it out. We want cloud to able to say that. And so these are some of the things will include in this final reminder and kind of a wrap up step for clud reminded to do things like answer only if it's very confident we could even ask it to refer back to what it has seen in the form any time, it's making a factual claim. So if it wants to say vehicle be turned, right, it should say, i know this. It's on a fact that box, too, is clearly checked or whatever it might be, we can kind of give clothds and guidelines about that. So if we go back to the console, we can see the next version of the prompts, and we're going to keep we're going to keep everything the same here in the system prompts. We're not changing any of that background context that we gave delaud about the form about how it's going to fill everything out. We're not changing anything else about the context in the rule. We're just adding some detailed list of text. And this is how we want laud to go about analyzing this and a really key thing that we found here. And we were building the demo. And when we were working on the customer example, is that the order in which laud analyzes this information is very important. And this is analogous to way you might think about doing this. If you were a human, you would probably not look at the drawing first and try to understand what was going on, right? It's pretty unclear. It's a bunch of boxes and lions. And we really know what that drawing st st of mean without it any additional of text. But if we have the form and we can read the form m first and understand that we're talking about a car accident and that we're seeing some checkboxes that indicate what vehicles were doing at certain times, then we is a little little it more about how to understand what might be in the drawing. And so that's the kind of detail that we're going to give cud here is to say, hey, first, go look at the form. Look it very carefully, make sure you can tell what boxes are checked, make sure you're not missing anything here. Make a list for yourself of what you see in mad and then move on to the sketch. So after youve kind of confidently gotten information out of the form, and you can say what's factually true, then you can go on and think about what you can gain in from that sketch, keeping in mind you're understanding of the accident so far. So whatever you've learned from the form, you're try to match that up with the sketch. And that's how you're going to arrive at your final at your final assessment of the form and will run it. And here you can see one behavior that this produced for cloud, because i told it to very carefully examine the form. It's showing me it's work as it does that. So it's telling me each individual box is the box checked. Is it not checked? And so this is one thing you'll notice, as you do prompt engineering in our previous prompse, we were kind of letting clow decide how much wanted to tell about about what it's so on the form here, because because i've told it carefully examine each and every box, it's very carefully examining each of every box. And that might not be what we want in the end. So that's something we might change. But it's also gonna give me in these other things that i ask for in examl tags. So a nice analysis of the form the accident summary so far, it's going to give me a sketchch analysis. And it's going to continue to say that vehicle be appear to clearly at fault in this exexample is prerace simple example, with more complicated drawings, more less clarity in the forms. This kind of step i step thinking for cloud is really impactful in its ability to make a correct assessment here. Uh, i think we'll go back to the slide. And going to talk about a last kind of piece that we might add to this to really make it useful for a real world task. D. Thank you so much. So as had i mentioned, we sort of set the stage in this problem to make sure that pulls are you acting on our behalf in a right manner? Um and it keesed up there were your frid towards the end of this problem. We're going to showing a second is a simple sort of guidelines of remind apart as well shyou're stggthening and reenforcing exactly what wanto get out of of it. And one important piece is actually output for wedding. You can imagine if you're a date engineer working on this olympplfication, all the sort of fancy preambo is great. But at another day, you want your piece information to to be stored in this, say, your sequal database whever. You want to sort a data. And the rest of it that it's necessary for cllad to sort of give its verdict isn't really unnecessary ily apppetition. You want the native beauty information for your application, safe一transition back to council. You'll see here that, which is added a simple importance guidelines part. And again, this is just reinforcing the sort of mechanical behavior. I won't undercloud here when make sure that the summaries clear conscience and accurate, want to make sure that nothing is sort of impeding in in enclose assessment, apart from the the as analyzing. And i finally, when it comes ulpera formatting, in my case here, i'm just going to ask called to wrap its final verdict all other stuff. I'm actually going ignore from my application and just look at what is actually sessing, and that is II can use this for going to build some sort of, and it it looks too often as well. Or if i just want to kar cut um determination, this is where i can do so. So i just run this here. You see you. It's going through the same sort processing seeing before, in this case is much more sucincct because i'have asked to be to summarize its findings and aumominstrafor d manner. And find me towards the end, you see that you wrap my output in these final verdict, external tags. So you see that during the stemo have gone from a skiing accidents,所以说unconfident in secure output from perhaps a car accident in the second version to know a much more strictly formatted confedent upera that we can tually build. And then like i might to application around and actually help no a real world um car insurance company, for example. Finally, if we transition back to the slides, another key way of shaping close output is actually putting words close mouth.我问，还是called precial responses, you could imagine that parsing exmell tags is nice and all, but maybe you want a structure, red jason output to make sure that uh, ststtions it's realizable, and you can use this in subsequent subsequent call. For example, um this is quite simple to do. You could just add that um claudds to begin its output with a certain four mat. This could be, for example, a uh open square back back squsquare ly back, for example, or even in this case that we seein front of us. This would be an exammultage fortunaary. In our case. It could also be that final verdict是茅台。 Um and this is just a great way of again, shaping how ploud supposed respond. Um without all the pre m with you don't want that, even though that is also key in shaping his output to make sure that called is reasoning through the steps that we wanted. So in our case here, we would just wrap it in the phaneotic and imparse it ofterwards, but you can use pief as well. Now finally, one step that would like a highhere as well as a both kills. You were seven and especially called four, of course, is as of hybrid reasoning model, meaning that there's extended thinking at your disposable. And this is something we want to highlight, because you can use extent thinking as a cruch for your prom engineering, basically can enable this to make sure that so actually has time to think it as this thinking tag and the scratch pad um and the beaudo it. That is you can actually analyze our transcaps understand how clooud is going about that data. So as we mentioned, we have these checheck boxes where goes to step a step of this aren't scenario that trspared the accidents. And in many ways, there you can actually try to help ploud and building this into rest system from himself. Is there only more token efficient? But it's a good way of understanding how these intelligent models that don't have oriantuition and actually go about the data that we cry them. And because that is quite key, i should turn to break down how your system prompt to get a lot better. Um i'm with that said, i think yeah like to thank all you're coming today. We will be around as well. If any questions tions on proproting, please please please ahead. I know there's a prompting you want to learn more about prompting in an hour. We have prompting for agents. And right now, we have an amazing demo of caldd place. Pokemon. So don't go any arewhere for that. And as christian said, it will be around all day. So i know me to have time for q and a in the sesbut, please confine us if you want to chat. And thank you. You guys for coming,还有什么可以？