Prompt Engineering 101: Best Practices for Analyzing Car Accident Claims

Hi everyone. Thank you for joining us today for  prompting 101. Uh my name is Hannah. I'm part   of the applied AI team here at Anthropic. And  with me is Christian, also part of the applied   AI team. And what we're going to do today is  take you through a little bit of prompting best  

practices. And we're going to use a real world  scenario and build up a prompt together. Uh so   a little bit about what prompt engineering is. uh  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 might want to think  about building out a prompt. Um, and as always,   the best way to learn this is just to practice  doing it. Um, so today we're going to go through   a hands-on scenario. Uh, we're going to use an  example inspired by a real customer that we worked  

with. So, we've modified what the actual customer  asked us to do, but this is a really interesting   case of trying to analyze some images and get uh  factual information out of the images and have   Claude make a judgment about what content it finds  there. And I actually do not speak the language  

that this content is in, but luckily Christian and  Claude both do. Uh so I'm going to pass it over   to Christian to talk about the scenario and the  content. So for this example that we have here,   it's uh intended so so to set the stage, imagine  you're working for a Swedish insurance company  

and you deal with uh car insurance claims on a  daily manner. Um and the purpose of this is that   you have two pieces of information. Um we're going  to these in detail as well, but visually you can   see on the left hand side we have a car accident  report form. um just detailing out what transpired  

before the action accident actually took place.  And then finally, we have a sort of human drawn   um sketch of how the accident took place as well.  So these two pieces of information is what we're   going to try to pass on to cloud. And to begin  with, we could just take these two and throw them  

into a console and just see what what happens.  So if we transition over to console as well,   we can actually do this in a real manner. And  in this case here, you can see we have our   shiny beautiful entropic console. We're using the  new claw for solid model as well. In this case,  

setting temperature zero and having a a huge max  token budget as well. Just helping us make sure   that there's no limitations to what CL can do. In  this case, you can see I have a very simple prompt   just setting the stage of what Cloud's supposed  to do. in this case mentioning that this is  

um intend to review a an accident report form uh  and eventually also determine um what happened   in an accident and who's at fault. So you can  see here with this very simple prompt if I just   run this let me go to preview. Uh we can see here  that Claude thinks that this is in relation to a  

skiing accident that happened on a street called  Chappangan. It's a very common street in Sweden.   Um and in many ways you can sort of understand  this innocent mistake in the sense that in our   prompt we actually haven't done anything to set  the stage on what is actually taking place here.  

So this sort of first guess is not too bad but  we still notice a lot of intuition that we can   bake into cloud. So if we switch back to the  slides you can see here that um in many ways   prompt engineering is a very iterative empirical  science. Uh in this case here, we could almost  

have a test case where Claude is supposed to make  sure it understands it's in a car or vehicular   environment, nothing to do with skiing. Uh and in  that way, you iteratively build upon your prompt   to make sure it's actually tackling the problem  you're intending to solve. Um and to do so,  

we'll go through some best practices of how we we  at Anthropic break this down internally and how we   recommend others to do so as well. So, we're going  to talk about some best practices for developing   a great prompt. Uh, first 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 chatbot with Claude going back   and forth having a more kind of conversational  style interaction. 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 Claude  

and have it nail the task the first time around  without needing to uh kind of move back and forth.   Uh, so the kind of structure that we recommend is  setting the task description up front. So telling   Claude, "What are you here 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 was showing,   the form and the drawing of the accident and how  they occurred. That's our dynamic content. This   might also be something you're retrieving from  another system, depending on what your use case  

is. We're going to give some detailed instructions  to Claude, so almost like a step-by-step list of   how we want Claude to go through the task and how  we want it to um tackle the reasoning. We may give   some examples to Claude. Here's an example of some  piece of content you might receive. Here's how you  

should respond when given that content. And at  the end, we usually recommend repeating anything   that's really important for Claude to understand  about this task. Kind of uh reviewing the   information with Claude, emphasizing things that  are extra critical and then telling Claude, "Okay,  

go ahead and do your work." So, here's another  view. This has a little bit more detail, a little   bit more of a breakdown, and we're going to walk  through each of these 10 points individually and   show you how we build this up, um, in the console.  So, the first couple things, um, Christian's  

going to talk about the task context and the tone  context. Perfect. So, yeah, if we begin with the   task context, as you realized when I went through  a little demo there, um, we didn't have much   elaborating what what scenario Chlo was actually  working within. And because of that, you can also  

tell that claw doesn't necessarily need to guess a  lot more on what you actually want from it. So in   our case, we really want to break that down, make  sure we can give more clear-cut instructions. Um,   and also make sure we understand what's the  task that we're asking Claw to do. Um, secondly,  

as well, we also make sure we add a little bit of  tone into it all. Um, key thing here is we want   Claw to stay factual and to stay confident. So if  uh, Claw can't understand what it's looking at,   we don't want to guess and just sort of mislead  us. We want to make sure that any assessment  

and in our case we want to make sure that we can  understand who's at fault here. We want to make   sure that assessment is as clear and as confident  as possible. If not, we're sort of losing track of   what we're doing. So if we transition back to the  the console, um we can jump to a V2 that we have  

here. So I'll just navigate to V2. And you can see  here um I'll also just illustrate the data because   we didn't really do that last time around just  to really highlight what we're looking at. So,   what we're seeing here, this is the car accident  report form, and it's just 17 different checkboxes  

going through what actually happened. You  can see there's a vehicle A and vehicle B,   both on the left and right hand side. And the main  purpose of this is that we want to make sure that   Claude can understand this manually generated data  to assess what's actually going on. And that is  

uh corroborated by if I navigate back here to this  sketch that we can highlight here as well. In this   case, the form is just a different um data point  for the same scenario. Um and in this case here, I   want to bake in more information into our version  two. Uh and by doing so, I'm actually elaborating  

a lot more on what's going on. So, you can see  here I'm specifying that uh this AI assistant is   supposed to help a human's claim claims adjuster  that's reviewing car accident report forms in   Swedish as well. Um, you can see here we're also  elaborating that it's a human-driven sketch of  

the incident and that you should not um make an  assessment if it's not actually fully confident.   And that's really key because if we run this,  you'll see that and you can see it's the same   settings as well. Clo my new shiny model zero  temperature as well. If we run this, we can see  

here what actually happens in this case. Um, CL is  able to pick up that uh now it's relating to car   accidents, not skiing accidents, which is great.  We can see it's able to pick up that vehicle A was   marked on on checkbox one and then vehicle B was  on 12. Um, and if we scroll down though, we can  

still tell that there's some information missing  for claw to make a fully confident determination   of who's at fault here. And this is great. This  is pertaining to a task set. Make sure you don't   make anything any claims that aren't um uh factual  and make sure you you only sort of assess things  

when you're when you're confident. But there's  a lot of information we're still missing here.   um regarding the form uh what the form actually  entails and a lot of that information is what   we want to want to bake into this LM application  as well and the best way of doing so is actually  

adding it to the system prompt which Hannah will  elaborate on. Um so back in the slides uh we have   the next item we're going to add to 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 to Claude to tell  Claude, here's the structure of the form you'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  going to change. And so this is a great type of   um information to put into the system prompt. Also  a great thing to use prompt caching for if you're   considering using prompt caching. This will always  be the same. And what this will help Claude 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 do a better job of reading the form  because it already knows um what to expect there.   So another thing I want to touch on here is how we  like to organize information in prompts. So Claude  

really loves structure, loves organization.  That's why we recommend following kind of a   standard structure in your prompts. And there's  a couple other tools you can use to help Claude   understand the information better. I also just  want to mention all of this is in our docs 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. Um anyway the uh  so some things you can use delimiters like XML  

tags also markdown is pretty useful to Claude  but XML tags are nice because you can actually   specify what's inside those tags. So we can tell  Claude here's here's user preferences. Now you're   going to read some content and these XML tags are  letting you know that everything wrapped in those  

tags is related to the user's preferences and  it helps Claude refer back to that information   maybe at later points in the prompt. Um, so I  want to show in the back in the console how we   actually do this in this case. And Christian's  going to pull 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 the system prompt.  You could try this different ways. Uh, we're doing   it in the system prompt here. And we're going  to tell Claude everything it needs to know about  

this form. So this is a Swedish car accident  form. The form will be in Swedish. It'll have   this title. It'll have two columns. The columns  represent different vehicles. We'll tell Claude   about each of the 17 rows and what they mean.  You might have noticed when we ran it before,  

Claude 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 Claude a little bit of information about   how this form should be filled out. This is also  really useful for Claude. We can tell it things  

like, you know, humans are filling this form  out basically. So, it's not going to be perfect.   People might put a circle. They might scribble.  They might not put an X in the box. There could   be many types of markings that you need to look  for when you're reading this form. Uh we can also  

give Claude a little bit of 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 going to help Claude   um 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.  Oh, your scroll is backwards from mine. Uh,   the we have the same user prompt here. Still  asking Claude to do the same task, same context.   And we'll see here that it's spending less time.  It's kind of narrating to us a little bit 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 whole list of what it found to be   checked, what the sketch shows. And here Claude  is now becoming much more confident with this  

additional context that we gave to Claude. Claude  now feels it's appropriate to say vehicle B was   at fault in this case based on this drawing and  based on this sketch. So already we're seeing some   improvement in the way Claude is analyzing these.  I think we could probably all agree if we looked  

at the drawing and at the list that vehicle  B is at fault. Um so we'd like to see that.   Uh so we're going to go back to the slides and  talk about a couple of other items that we're not   really using in this prompt um but can be really  helpful to building up uh building up your prompt  

and making it work better. Exactly. I think um  one thing that we really highlight is examples.   I think examples or few shot is a mechanism that  really is powerful in steering cloud. So you can   imagine this um in in quite a non-trivial way as  well. So imagine you have scenarios, situations,  

even in this case concrete accidents that have  happened that are um tricky for claw to get right.   But you with your human intuition and your human  label data um is able to actually get to the right   conclusion. Then you can bake that information  into the system problem itself by having clear-cut  

examples of a the data that that it's supposed  to look at. So you can have visual examples.   you can just base 64 encode a a an image and have  that as part of the data that you're passing along   into the examples and then on top of that you can  have the sort of depiction or description rather  

of how to break that down and understand it. This  is something we really highlight and and emphasize   in how you can sort of push the limits of your  LLM application is by baking in these examples   into system prompt. And this again is sort of the  empirical science of prompt engineering that you  

sort of always want to push the limits of your  application and get that feedback loop in where   it's going wrong and try to add that into system  prompt so that next time when example that sort   of mimics that u takes place it's able to actually  reference it in its example set. You can see here  

as well, this is just a little example of how we  do this. Again, really emphasizing the sort of XML   structure that we we um we enjoy. It's it gives a  lot of structure to the clone. It's what it's been   fine-tuned on as well. Um and it works perfectly  well for this example. And in our case, we're not  

doing this just because it's a simple demo,  but you can realistically imagine if you were   building this for an insurance company, you'd have  tens, maybe even hundreds of examples are quite   difficult, maybe in the gray, that you'd like to  make sure that Claude actually has some basis in  

to make the verdict next time. Um, another topic  we really want to highlight, which we're not doing   in this demo, is conversation history. It's in the  same vein as examples. uh we use this to make sure   that the enough context rich information is at  close disposal when it when when closing on on on  

your behalf. Um in our case now this isn't really  a userfacing LLM application. It's more something   happening in the background. You can imagine for  this insurance company they have this automated   system some data is generated out of this and then  you might have a human in the loop at towards the  

end. If you were have to build something much more  userf facing where you'd have a long conversation   history that would be um relevant to bring in this  is a perfect place in the system prompt to include   because it enriches the context that Claude 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 task at hand.   So, now we're going to build out the  final part of this prompt for Claude,   and that's coming back to the reminder of what  the immediate task is and giving Claude a reminder  

about any important guidelines that we want it  to follow. Some reasons that we may do this are   a preventing hallucinations. Um, so we want Claude  to uh not invent details that it's not finding in   this prompt, right? Or not finding in the data.  If Claude can't tell which form is checked,  

we don't want Claude to take its best guess or  invent the idea that a box might be checked when   it's not. If the sketch is unintelligible, the  person did a really bad job drawing this drawing   and even a human would not be able to figure it  out. We want Claude to be able to say that. And  

so these are some of the things we'll include in  this final reminder and kind of wrap up step for   Claude. Uh remind it 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 anytime  it's making a factual claim. So if it wants to say  

vehicle B turned right, it should say I know this  based on the fact that box two is clearly checked   or whatever it might be. We can kind of give  Claude some guidelines about that. So if we go   back to the console, we can see the next version  of the prompt and we're going to keep uh we're  

going to keep everything the same here in the  system prompt. So, we're not changing any of that   background context that we gave to Claude about  the form, about how it's going to fill everything   out. We're not changing anything else about the  context and the role. We're just adding this  

detailed list of tasks. And this is how we want  Claude to go about analyzing this. And a really   key thing that we found here as we were building  this demo and when we were working on the customer   example is that the order in which Claude 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 lines. We don't really know  what that drawing is supposed to mean without any  

additional context. But if we have the form and we  can read the form first and understand that we're   talking about a car accident and that we're seeing  some checkboxes that indicate what vehicles we're   doing at certain times, then we know a little  bit 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 Claude here is to say,   "Hey, first go look at the form. Look at it very  carefully. Make sure you can tell what boxes are   checked. Make sure you're not missing anything  here. Um, make a list for yourself of what you see  

in that. And then move on to the sketch. So after  you've 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 from that sketch. keeping in mind your  understanding of the accident so far. So, whatever  

you've learned from the form, you're trying to  match that up with the sketch. And that's how   you're going to arrive um at your final uh at your  final assessment of the form. And we'll run it. And here you can see one behavior that  this produced for Claude because I told  

it to very carefully examine the form. It's  showing me its 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 prompts, we were kind of letting  

claw decide how much it wanted to tell us about  what it saw on the form here. Because I've told   it carefully examine each and every box, it's very  carefully examining each and every box. And that   might not be what we want in the end. So, that's  something we might change. Um, but it's also going  

to give me these other things that I asked for  in XML tags. So, a nice analysis of the form, the   accident summary so far. It's going to give me a  sketch analysis, and it's going to continue to say   that vehicle B appears to be clearly at fault. In  this in this example, it's pretty simple example  

with more complicated drawings, more uh less  clarity in the forms. This kind of step-by-step   thinking for Claude is really impactful in  its ability to make a correct assessment here.   Uh, so I think we'll go back to the slides and  Christian's going to talk about a last kind of  

piece that we might add to this um to really make  it useful for a real world task. Indeed. Thank   you so much. So, as Hannah mentioned, uh, we sort  of set the stage in this prompt to make sure that   really acting on our behalf in a right manner.  Um, and a key step that we also add towards the  

end of this prompt that I'm going to show you in a  second is a simple sort of guidelines or reminder   part as well. just strengthening and reinforcing  exactly what we want to get out of it. And one   important piece is actually output formatting.  You can imagine if you're a data engineer working  

on this LM application, all the sort of fancy  preamble is great, but at the end of the day,   you want your piece of information to to  be stored in, let's say, your SQL database,   wherever you want to store that data. And the rest  of it that is necessary for cloud to sort of give  

its verdict isn't really that necessary for your  application. You want the nitty-gritty information   for your application. So if we transition back to  the console, you'll see here that we just added   a simple importance guidelines part. And again,  this is just reinforcing the sort of mechanical  

behavior that we want out of cloud here. Want  to make sure that the summary is clear, concise,   and accurate. Want to make sure that nothing  is sort of impeding in in in Claw's assessment   apart from the data it's analyzing. And then  finally, when it comes to output formatting,  

in my case here, I'm just going to ask Claude  to wrap its final verdict. All other stuff I'm   actually going to ignore for my application and  just look at what it's actually assessing. And   that is I can I can use this if I want to build  some sort of analytics tool afterwards as well.  

Or if I just want a clearcut um uh determination,  this is a way I can do so. So if I just run this   here, you'll see it's going through the same sort  of process that we've seen before. In this case,   it's much more succinct because we've asked  to be to summarize its findings in a much  

more straightforward manner. And then finally  towards the end you'll see that it'll wrap my   output in these final verdict XML tags. So you  can see that during this demo we've gone from   a skiing accident to sort of unconfident insecure  outputs from perhaps a car accident in the second  

version to now a much more strictly formatted  confident output that we can actually build an   application around and actually help you know a  real world um car insurance company for example.   U finally if we transition back to the um slides  another key way of shaping CL's output is actually  

putting words in CL's mouth or as we call it  pre-filled responses. You could imagine that   parsing XML tags is nice and all but maybe you  want a structured JSON output to make sure that   uh it's JSON serializable and you can use this  in a subse subsequent call for example. Um this  

is quite simple to do. You could just add that um  claude needs to begin its output with a certain   format. This could be for example a uh open  square bracket squarely bracket for example   or even in this case that we see in front of us  this would be an XML tag for itinerary. In our  

case it could also be that final verdict XML tag.  Um, and this is just a great way of again shaping   how Claude is supposed to respond. Um, without all  the preamble if you don't want that, even though   that is also key in shaping his output to make  sure that Claude is reasoning through the steps  

that we wanted. So in our case here, we would just  wrap it in the final verdict and then parse it   afterwards. But you can use prefill as well. Now  finally one step that I would like to highlight   here as well is that both cloud 3.7 and especially  cloud 4 of course is a sort of hybrid reasoning  

model meaning that there's extended thinking at  your disposal. Um and this is something we want   to highlight because you can use extended thinking  as a crutch for your prompt engineering. Basically   you can enable this to make sure that Claude  actually has time to think. It adds his thinking  

tags and the scratch pad. Um and the beauty of  that is you can actually analyze that transcript   to understand how claude is going about that data.  So as we mentioned we have these check boxes where   it goes through step by step of the scenario  that transpired for the accident. And in many  

ways there you can actually try to help claude in  building this into the system prompt itself. It's   not only more token efficient but it's a good way  of understanding how these intelligent models that   don't have our intuition actually go about the  data that we provide them. And because of that,  

it's quite key in actually trying to break down  how your system prompt can get a lot better. Um,   and with that said, I think uh I'd like to thank  all you for coming today. We'll be around as well.   So if you have any questions on prompting, please  uh please go 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 Claude plays Pokemon. So  don't go anywhere for that. And as Christian said,   we'll be around all day. So, I know we  didn't have time for Q&A in this session,  

but uh please come find us if you want to chat.  And thank you guys for coming. Thank you so much.