SUMM

The presenters introduce themselves as part of Anthropic's applied AI team and set the agenda to cover prompt engineering best practices using a real-world example.
The session will involve building a prompt to analyze images and make factual judgments, specifically for a Swedish car insurance claim scenario.
The scenario involves a car accident report form and a human-drawn sketch, both of which Claude (the AI model) will analyze to determine fault.

The first attempt at prompting Claude is minimal, resulting in incorrect context detection (Claude assumes a skiing accident instead of a car accident).
The initial lack of explicit scenario context leads Claude to make reasonable but incorrect guesses.
Prompt engineering is highlighted as iterative and empirical—iterations are necessary to refine the prompt based on observed outputs.

Best practices for prompt structure include:
- Setting a clear task description.
- Providing relevant content (e.g., forms, images).
- Supplying detailed step-by-step instructions.
- Including examples when possible.
- Concluding with a reminder or summary of key points.
When working with APIs, a single well-crafted prompt is preferable to conversational back-and-forth.

Adding explicit task context helps guide the AI to the right problem space (e.g., specifying car accident, not skiing).
Including tone instructions encourages factual, confident assessments and discourages guessing when information is unclear.
Showing how revised prompts with clearer context lead Claude to more relevant responses, but still with some information missing for full confidence.

Details that do not change (like the structure of the accident form) are recommended to be placed in the system prompt for efficiency and clarity.
Using delimiters like XML tags or Markdown helps structure the information for Claude; XML is particularly favored for clarity and reference.
Organization within the prompt aids Claude’s comprehension, leading to more accurate and confident outputs.

Including specific examples (few-shot learning) in the prompt/system prompt can significantly improve Claude’s accuracy, especially for tricky or ambiguous cases.
Examples can be textual descriptions or encoded images, coupled with the expected analysis or conclusion.
Reiterates the empirical, iterative approach to refining prompts by identifying where the model struggles and incorporating relevant examples.

When relevant, past conversation history can be included in prompts to provide richer context, though not needed for the current backend-focused demo.
Explicit reminders and guidelines at the end of the prompt help prevent hallucination and ensure Claude only answers when confident.
Guidelines can instruct Claude to cite evidence directly from the input data for any factual claims.

Detailed step-by-step task lists improve Claude’s reasoning by specifying the order and method of analysis (e.g., examine the form first, then the sketch).
The prompt can be tailored to produce outputs in desired structures (e.g., XML tags, JSON), aiding downstream processing and integration with other systems.
By iterating on task instructions and output format requirements, the model output becomes more succinct, accurate, and directly usable.

Output formatting instructions (like wrapping verdicts in specified tags) are essential for integration and for filtering only required results.
Prefilled response structures (e.g., beginning output with a certain tag or JSON format) further ensure model compliance with required formats.
Advanced model versions (like Claude 3.7 and 4) support extended thinking modes, allowing the model to show its reasoning in scratchpads—helpful for both engineering and debugging.

Emphasis on the continuous, iterative nature of prompt engineering and the value of structured, explicit communication with the AI model.
Attendees are invited to further sessions and live demos, as well as to engage with the presenters for questions outside of the recorded session.

Prompting 101