Rahul · Launch Video Breakdown: Hook, Pacing & Motion Design

Scene-by-scene timeline & spoken transcript

1. 0:00 – 4:56

   The Hook

   Introduction to Prompt Engineering & Claude's Capabilities

   “What is prompt engineering anyway? The practice of systematically improving prompts for LLM applications through testing, evaluation, analysis, and optimization of prompts. Skills involved: Programming in natural language, clear, unambiguous, precise writing, conceptual engineering, creating evals with a scientific mindset, product thinking - what is the ideal model behavior for your product?, testing, understanding the LLMs, aggregating and analyzing failure modes - thinking of ways to fix, making LLMs scale to a wide range of inputs. Let's build a great prompt from scratch in the console. 1. Task context. You are an AI assistant helping a human claims adjuster review car accident reporting documentation written in Swedish. Your task is to review a standard accident report form as well as human-drawn sketch of the incident, and determine what information you can confidently conclude from these documents. After that, you should be able to state whether or not you can fully confidently determine that one vehicle was clearly at fault or if the human adjuster needs to follow up for more information. Based on my review of these Swedish accident report documents, here's what I can confidently conclude: Information from the Documents: From the Standard Form: This is section 12 of a standard accident report asking "How did the accident happen? (Hur kom det sig att olyckan hände?)" Vehicle A has marked option 1: "parkerade / stannade till" (was parking/stopping) Vehicle B has marked option 12: "tog av till höger" (turned off to the right) From the Sketch: Shows two vehicles (A and B) at what appears to be an intersection Vehicle A is marked as stationary ("still stående") The street is labeled "KÖPMANGATAN" Vehicle B appears to be making a turning movement. Add to Conversation. Let's build a great prompt from scratch in the console. 1. Task context. 2. Tone context. 3. Background data, documents, and images. 4. Detailed task description & rules. 5. Examples. 6. Conversation history. 7. Immediate task description or request. 8. Thinking step by step / take a deep breath. 9. Output formatting. 10. Prefilled response (if any).”

   On screen

   What is 'prompt engineering' anyway? The practice of systematically improving prompts for LLM applications through testing, evaluation, analysis, and optimization of prompts. Skills involved: Programming in natural language Clear, unambiguous, precise writing Conceptual engineering Creating evals with a scientific mindset Product thinking - what is the ideal model behavior for your product? Testing Understanding the LLMs Aggregating and analyzing failure modes - thinking of ways to fix Making LLMs scale to a wide range of inputs ANTHROPIC Console Workbench Batches System Prompt Define a role, tone or context (optional) User You are an AI assistant helping a human claims adjuster review car accident reporting documentation written in Swedish. Your task is to review a standard accident report form as well as human-drawn sketch of the incident, and determine what information you can confidently conclude from these documents. After that, you should be able to state whether or not you can fully confidently determine that one vehicle was clearly at fault or if the human adjuster needs to follow up for more information. form-numbered.jpg IMG_8157.jpg Pre-fill response Add message pair Prompt Evaluate Get Code Run + Templatize Response Preview API type: "message", "role": "assistant", "content": [ { "type": "text", "text": "Based on my review of this Swedish accident report form and diagram, here's what happened:\n\nAccording to the form selections:\n1. Person A was stationary/standing still (\"parkerade/stannade till\") - marked with a checkmark\n2. Person B was turning to the right (\"tog av till höger\") - marked with a checkmark\n\nThe diagram shows Person A was stationary on the Add to Conversation Christian Ryan Anthropic (Normal Acc...) Prompt structure 1-2 sentences to establish role and high level task description 2 Dynamic / retrieved content 3 Detailed task instructions 4 Examples-to-shot (optional) 5 Repeat critical instructions (particularly for very long prompts) Let's build a great prompt from scratch in the console 1. Task context ANTHROPIC Console Workbench Batches System Prompt Define a role, tone or context (optional) User You are an AI assistant helping a human claims adjuster review car accident reporting documentation written in Swedish. Your task is to review a standard accident report form as well as human-drawn sketch of the incident, and determine what information you can confidently conclude from these documents. After that, you should be able to state whether or not you can fully confidently determine that one vehicle was clearly at fault or if the human adjuster needs to follow up for more information. form-numbered.jpg IMG_8157.jpg Pre-fill response Add message pair Prompt Evaluate Get Code Run + Templatize Response Preview API Based on my review of these Swedish accident report documents, here's what I can confidently conclude: Information from the Documents: From the Standard Form: This is section 12 of a standard accident report asking "How did the accident happen? (Hur kom det sig att olyckan hände?)" Vehicle A has marked option 1: "parkerade / stannade till" (was parking/stopping) Vehicle B has marked option 12: "tog av till höger" (turned off to the right) From the Sketch: Shows two vehicles (A and B) at what appears to be an intersection Vehicle A is marked as stationary ("still stående") The street is labeled "KÖPMANGATAN" Vehicle B appears to be making a turning movement. Add to Conversation Christian Ryan Anthropic (Normal Acc...) ANTHROPIC Console Workbench Batches System Prompt Define a role, tone or context (optional) User You are an AI assistant helping a human claims adjuster review car accident reporting documentation written in Swedish. Your task is to review a standard accident report form as well as human-drawn sketch of the incident, and determine what information you can confidently conclude from these documents. After that, you should be able to state whether or not you can fully confidently determine that one vehicle was clearly at fault or if the human adjuster needs to follow up for more information. Here is the form: <form_image> form-numbered.jpg </form_image> Pre-fill response Add message pair Prompt Evaluate Get Code Run + Templatize Response Preview API <sketch_interpretation> stationary Vehicle A Conclusion I can confidently determine fault in this accident. Vehicle B is clearly at fault because: 1. Vehicle A was parked/stopped (stationary) as indicated by the checked box and confirmed by the sketch notation "still stående" 2. Vehicle B was actively turning right when the collision occurred 3. A moving vehicle that strikes a stationary vehicle is typically at fault, especially when the stationary vehicle was properly parked/stopped Add to Conversation Christian Ryan Anthropic (Normal Acc...) Let's build a great prompt from scratch in the console 1. Task context 2. Tone context 3. Background data, documents, and images 4. Detailed task description & rules 5. Examples 6. Conversation history 7. Immediate task description or request 8. Thinking step by step / take a deep breath 9. Output formatting 10. Prefilled response (if any)

   Camera

   Alternates between wide shots of the presenters and audience, close-ups of the presenters, and screen recordings of the Anthropic console. The camera occasionally zooms in slightly on the presenters.

   Motion

   Seamless transitions between live presentation footage and screen recordings. Text highlights and cursor movements within the console are animated to draw attention to key features. On-screen text elements appear with a subtle fade-in or slide-in effect.

2. 4:56 – 9:52

   Problem Agitation

   Refining Prompts for Accuracy and Confidence

   “So we're going to start with a very simple prompt. We're going to add a little bit of context. We're going to add some detailed instructions. We're going to add some examples. We're going to add some conversation history. We're going to add an immediate task description. We're going to add some thinking step by step. We're going to add some output formatting. And we're going to add some pre-filled response. And we're going to see how that changes the output of Claude. So let's start with a very simple prompt. You are an AI assistant helping a human claims adjuster review car accident reporting documentation written in Swedish. Your task is to review a standard accident report form as well as human-drawn sketch of the incident, and determine what information you can confidently conclude from these documents. After that, you should be able to state whether or not you can fully confidently determine that one vehicle was clearly at fault or if the human adjuster needs to follow up for more information. So this is our initial prompt. And we're going to run this. And we're going to see what Claude says. So Claude says, based on my review of these Swedish accident report documents, here's what happened. According to the form selections, Person A was stationary/standing still, marked with a checkmark. Person B was turning to the right, marked with a checkmark. The diagram shows Person A was stationary on the street labeled KÖPMANGATAN. Person B appears to be making a turning movement. So Claude is doing a pretty good job of summarizing the information. But it's not making a clear determination of fault. So we want to add some more context to this prompt. So we're going to add some tone context. We're going to say, you are an AI assistant helping a human claims adjuster review car accident reporting documentation written in Swedish. Your task is to review a standard accident report form as well as human-drawn sketch of the incident, and determine what information you can confidently conclude from these documents. After that, you should be able to state whether or not you can fully confidently determine that one vehicle was clearly at fault or if the human adjuster needs to follow up for more information. And we're going to add some background data, documents, and images. So we're going to add the form-numbered.jpg and IMG_8157.jpg. And we're going to add some detailed task description and rules. So we're going to say, important guidelines: Base your analysis solely on the information visible in the images. Do not make assumptions or add details not explicitly shown. If certain information is unclear or not provided, state this in your summary. Ensure your summary is clear, concise, and accurate. If you cannot access or view the image for any reason, please state so immediately and do not proceed with the analysis. If you cannot make a clear determination about what boxes are checked or what numbers are written, please state so immediately and do not proceed with the analysis. Now, please start your step by step analysis. Wrap your final verdict in <final_verdict> XML tags. So we're going to run this again. And we're going to see what Claude says. So Claude says, examining the form carefully for Vehicle A (left column) and Vehicle B (right column): Vehicle A: Box 1: There is a clear diagonal line/mark across this checkbox, indicating it is checked. Boxes 2-17: All appear to be empty with no markings. Vehicle B (Right Column): Box 12: There appears to be a diagonal line/mark across this checkbox, indicating it is checked. Boxes 1-11, 13-17: All appear to be empty with no markings. Bottom Section: The total count boxes at the bottom appear to have some markings but are not clearly legible in this image quality. So Claude is doing a much better job of breaking down the information. But it's still not making a clear determination of fault. So we want to add some examples.”

   On screen

   ANTHROPIC Console Workbench Batches System Prompt Define a role, tone or context (optional) User You are an AI assistant helping a human claims adjuster review car accident reporting documentation written in Swedish. Your task is to review a standard accident report form as well as human-drawn sketch of the incident, and determine what information you can confidently conclude from these documents. After that, you should be able to state whether or not you can fully confidently determine that one vehicle was clearly at fault or if the human adjuster needs to follow up for more information. form-numbered.jpg IMG_8157.jpg Pre-fill response Add message pair Prompt Evaluate Get Code Run + Templatize Response Preview API type: "message", "role": "assistant", "content": [ { "type": "text", "text": "Based on my review of these Swedish accident report documents, here's what happened:\n\nAccording to the form selections:\n1. Person A was stationary/standing still (\"parkerade/stannade till\") - marked with a checkmark\n2. Person B was turning to the right (\"tog av till höger\") - marked with a checkmark\n\nThe diagram shows Person A was stationary on the street labeled KÖPMANGATAN. Person B appears to be making a turning movement." } ] Add to Conversation Christian Ryan Anthropic (Normal Acc...) Let's build a great prompt from scratch in the console 1. Task context 2. Tone context 3. Background data, documents, and images 4. Detailed task description & rules 5. Examples 6. Conversation history 7. Immediate task description or request 8. Thinking step by step / take a deep breath 9. Output formatting 10. Prefilled response (if any) ANTHROPIC Console Workbench Batches System Prompt Define a role, tone or context (optional) User You are an AI assistant helping a human claims adjuster review car accident reporting documentation written in Swedish. Your task is to review a standard accident report form as well as human-drawn sketch of the incident, and determine what information you can confidently conclude from these documents. After that, you should be able to state whether or not you can fully confidently determine that one vehicle was clearly at fault or if the human adjuster needs to follow up for more information. Here is the form: <form_image> form-numbered.jpg </form_image> Pre-fill response Add message pair Prompt Evaluate Get Code Run + Templatize Response Preview API <tasks id="4"> Important guidelines: - Base your analysis solely on the information visible in the images. Do not make assumptions or add details not explicitly shown. - If certain information is unclear or not provided, state this in your summary. - Ensure your summary is clear, concise, and accurate. - If you cannot access or view the image for any reason, please state so immediately and do not proceed with the analysis. - If you cannot make a clear determination about what boxes are checked or what numbers are written, please state so immediately and do not proceed with the analysis. - Now, please start your step by step analysis. - Wrap your final verdict in <final_verdict> XML tags. Examining the form carefully for Vehicle A (left column) and Vehicle B (right column): Vehicle A: Box 1: There is a clear diagonal line/mark across this checkbox, indicating it is checked. Boxes 2-17: All appear to be empty with no markings. Vehicle B (Right Column): Box 12: There appears to be a diagonal line/mark across this checkbox, indicating it is checked. Boxes 1-11, 13-17: All appear to be empty with no markings. Bottom Section: The total count boxes at the bottom appear to have some markings but are not clearly legible in this image quality. Add to Conversation Christian Ryan Anthropic (Normal Acc...)

   Camera

   Similar to the previous beat, with frequent cuts between the presenters and the Anthropic console. The console view often highlights specific text inputs or outputs as the speakers discuss prompt refinements.

   Motion

   Animated text highlighting and cursor movements within the console. On-screen bullet points and numbered lists animate in sequentially to match the spoken dialogue. Subtle camera pushes into the console screen to emphasize details.

3. 9:52 – 19:48

   Product Reveal

   Advanced Prompt Engineering Techniques with Claude

   “So we're going to add some examples. And we're going to add some conversation history. And we're going to add an immediate task description. And we're going to add some thinking step by step. And we're going to add some output formatting. And we're going to add some pre-filled response. And we're going to see how that changes the output of Claude. So let's add some examples. So we're going to add an example of a prompt where Claude correctly identifies fault. And we're going to add an example of a prompt where Claude correctly identifies that more information is needed. And we're going to add some conversation history. So we're going to add a previous turn where Claude asked for more information. And we're going to add an immediate task description. So we're going to say, now, please provide a final verdict. And we're going to add some thinking step by step. So we're going to say, think step by step. And we're going to add some output formatting. So we're going to say, wrap your final verdict in <final_verdict> XML tags. And we're going to add some pre-filled response. So we're going to pre-fill the response with a confident determination of fault. So we're going to run this again. And we're going to see what Claude says. So Claude says, examining the form carefully for Vehicle A (left column) and Vehicle B (right column): Vehicle A: Box 1: There is a clear diagonal line/mark across this checkbox, indicating it is checked. Boxes 2-17: All appear to be empty with no markings. Vehicle B (Right Column): Box 12: There appears to be a diagonal line/mark across this checkbox, indicating it is checked. Boxes 1-11, 13-17: All appear to be empty with no markings. Bottom Section: The total count boxes at the bottom appear to have some markings but are not clearly legible in this image quality. Thinking Process: 1. Analyze the form data for Vehicle A and Vehicle B. 2. Analyze the sketch for Vehicle A and Vehicle B. 3. Compare the form data and sketch to determine if there is a clear fault. 4. If a clear fault is determined, state it confidently. 5. If more information is needed, state that. Final Verdict: I can confidently determine fault in this accident. Vehicle B is clearly at fault because: 1. Vehicle A was parked/stopped (stationary) as indicated by the checked box and confirmed by the sketch notation "still stående" 2. Vehicle B was actively turning right when the collision occurred 3. A moving vehicle that strikes a stationary vehicle is typically at fault, especially when the stationary vehicle was properly parked/stopped So Claude is now making a clear determination of fault. And it's providing a clear explanation for why it's making that determination. So this is a much better response. And this is what we're trying to achieve with prompt engineering. So we're going to talk about some advanced techniques now. We're going to talk about using XML tags. We're going to talk about prompt caching. And we're going to talk about extended thinking. So let's start with XML tags. So XML tags are a way to structure your prompts so that Claude can better understand the different parts of your prompt. So for example, you can use <task> tags to define the task that you want Claude to perform. You can use <context> tags to provide context to Claude. You can use <instructions> tags to provide detailed instructions to Claude. And you can use <examples> tags to provide examples to Claude. So XML tags are a very powerful way to structure your prompts. And they can help Claude to better understand your prompts and to produce better responses. So let's talk about prompt caching. So prompt caching is a way to store your prompts so that you don't have to re-type them every time you want to use them. So for example, if you have a prompt that you use frequently, you can save it as a cached prompt. And then you can just load it whenever you want to use it. So prompt caching can save you a lot of time and effort. And it can help you to be more efficient with your prompt engineering. So let's talk about extended thinking. So extended thinking is a way to get Claude to think step by step through a problem. So for example, if you have a complex problem, you can ask Claude to think step by step through the problem. And Claude will then break down the problem into smaller steps and will provide a detailed explanation of how it's solving each step. So extended thinking can help Claude to solve complex problems and to produce more accurate responses. So these are some advanced techniques that you can use with Claude. And they can help you to be more effective with your prompt engineering. So we're going to open it up for Q&A now. And we're going to take some questions from the audience. And then we're going to do a demo where Claude plays Pokémon.”

   On screen

   ANTHROPIC Console Workbench Batches System Prompt Define a role, tone or context (optional) User You are an AI assistant helping a human claims adjuster review car accident reporting documentation written in Swedish. Your task is to review a standard accident report form as well as human-drawn sketch of the incident, and determine what information you can confidently conclude from these documents. After that, you should be able to state whether or not you can fully confidently determine that one vehicle was clearly at fault or if the human adjuster needs to follow up for more information. Here is the form: <form_image> form-numbered.jpg </form_image> Pre-fill response Add message pair Prompt Evaluate Get Code Run + Templatize Response Preview API <tasks id="4"> Important guidelines: - Base your analysis solely on the information visible in the images. Do not make assumptions or add details not explicitly shown. - If certain information is unclear or not provided, state this in your summary. - Ensure your summary is clear, concise, and accurate. - If you cannot access or view the image for any reason, please state so immediately and do not proceed with the analysis. - If you cannot make a clear determination about what boxes are checked or what numbers are written, please state so immediately and do not proceed with the analysis. - Now, please start your step by step analysis. - Wrap your final verdict in <final_verdict> XML tags. Examining the form carefully for Vehicle A (left column) and Vehicle B (right column): Vehicle A: Box 1: There is a clear diagonal line/mark across this checkbox, indicating it is checked. Boxes 2-17: All appear to be empty with no markings. Vehicle B (Right Column): Box 12: There appears to be a diagonal line/mark across this checkbox, indicating it is checked. Boxes 1-11, 13-17: All appear to be empty with no markings. Bottom Section: The total count boxes at the bottom appear to have some markings but are not clearly legible in this image quality. Thinking Process: 1. Analyze the form data for Vehicle A and Vehicle B. 2. Analyze the sketch for Vehicle A and Vehicle B. 3. Compare the form data and sketch to determine if there is a clear fault. 4. If a clear fault is determined, state it confidently. 5. If more information is needed, state that. Final Verdict: I can confidently determine fault in this accident. Vehicle B is clearly at fault because: 1. Vehicle A was parked/stopped (stationary) as indicated by the checked box and confirmed by the sketch notation "still stående" 2. Vehicle B was actively turning right when the collision occurred 3. A moving vehicle that strikes a stationary vehicle is typically at fault, especially when the stationary vehicle was properly parked/stopped Add to Conversation Christian Ryan Anthropic (Normal Acc...) Let's build a great prompt from scratch in the console 1. Task context 2. Tone context 3. Background data, documents, and images 4. Detailed task description & rules 5. Examples 6. Conversation history 7. Immediate task description or request 8. Thinking step by step / take a deep breath 9. Output formatting 10. Prefilled response (if any)

   Camera

   Continues to alternate between presenters and console views. The console view often zooms in on specific XML tags or sections of the prompt and response to highlight the impact of advanced techniques.

   Motion

   Animated text highlighting, cursor movements, and scrolling within the console. On-screen bullet points and numbered lists animate in sequentially. Subtle camera pushes and pulls to emphasize different parts of the console interface.

4. 19:48 – 24:47

   Feature Teaser

   Q&A and Future Demos

   “So we're going to open it up for Q&A now. And we're going to take some questions from the audience. And then we're going to do a demo where Claude plays Pokémon.”

   On screen

   (No on-screen text)

   Camera

   Primarily wide shots of the presenters and the audience, with occasional close-ups of the presenters as they interact with the audience. The camera remains relatively static during the Q&A portion.

   Motion

   Minimal motion graphics, focusing on the live interaction. Simple cuts between speakers and audience members asking questions. No on-screen text animations or console demonstrations.