Anthropic Prompt Engineering for Business Performance

TL;DR

Anthropic has released a guide on prompt engineering for business performance, demonstrating that strategic prompting can significantly increase the accuracy, consistency, and cost-efficiency of Claude. By applying structured techniques like few-shot prompting and prompt chaining, businesses can reduce hallucinations and lower deployment costs while maintaining brand-aligned customer experiences.

Core Benefits of Effective Prompting

Strategic prompt engineering is essential for optimizing generative AI outputs in a professional environment. Anthropic identifies four primary advantages of well-crafted prompts:

Accuracy: Effective prompting further reduces the risk of inaccurate outputs and hallucinations.
Consistency: Well-thought-out prompts ensure uniform quality, formatting, relevance, and tone across end-user experiences.
Usefulness: Prompting allows businesses to cater to specific personas and industry-specific needs, delivering targeted experiences.
Cost Savings: Optimizing prompts minimizes unnecessary back-and-forth interactions, reducing the cost of running inputs and outputs at scale.

Key Prompting Techniques for Business

Anthropic recommends treating prompt engineering as a science—testing and iterating often—and providing clear, explicit instructions as if Claude were an intern on their first day.

Step-by-Step Reasoning

Directing Claude to "think step by step" often increases response accuracy. This is particularly effective for logic puzzles or complex evaluations, such as determining if an insurance claim meets specific policy requirements.

Few-Shot Prompting

Providing realistic and specific examples of desired inputs and ideal outputs helps Claude understand the expected format and handle edge cases. For instance, when redacting personally identifiable information (PII), providing synthetic examples of how names, phone numbers, and emails should be replaced with "XXX" ensures the model adheres to the required redaction standard.

Prompt Chaining

Prompt chaining involves breaking a complex task into multiple smaller prompts. Each subsequent prompt can include previous prompt-response pairs to build context. An example workflow for explaining tax situations would involve:

Creating a list of related tax codes.
Identifying relevant sections within those documents.
Generating a final response to the user based on the gathered information.

Case Study: Improving Accuracy by 20%

Anthropic partnered with a Fortune 500 company to build a customer-facing chat assistant using Claude Instant. The company sought a friendly, concise chatbot with low latency and high accuracy, having found other market solutions too wordy or stilted.

Implementation Strategy

Anthropic prompt engineers collaborated with the company's subject matter experts (SMEs) to implement the following:

Scratchpads: Directing Claude to use a scratchpad (e.g., using <relevant_quotes> tags) to show its work internally. This process is hidden from the customer but improves accuracy.
Few-Shot Examples: Training the model on the company's specific desired format and style.
SME Guidance: Directing Claude to use specific data points and workflows recommended by subject matter experts, such as specific factors affecting legal situations.

Results

By combining prompt engineering best practices with subject matter expertise, the company improved Claude's accuracy by 20%, while simultaneously reducing costs and accelerating time-to-market.