Anthropic explains how prompt engineering can improve the performance of Claude-based enterprise solutions. In a published post dated February 29, 2024, the company describes a case where a Fortune 500 firm increased Claude Instant’s accuracy by 20% through prompt engineering and collaboration between prompt engineers and subject-matter experts. The article also lists practical prompting techniques that reduce hallucinations, improve consistency, and lower operating costs.
Why prompt engineering matters
As businesses build on generative AI models, carefully constructed prompts are critical to producing high-quality outputs. Anthropic argues that the right prompts let organizations extract more value from Claude across tasks such as data processing, customer question answering, and contract review, while keeping outputs useful, relevant, and aligned with corporate standards and tone.
Benefits highlighted include:
- Accuracy: Prompt engineering further reduces the risk of inaccurate outputs, complementing other work to lower Claude’s hallucination rates.
- Consistency: Well-designed prompts ensure uniform quality, formatting, relevance, and tone for end users.
- Usefulness: Targeted prompts enable domain- and persona-specific experiences.
- Cost savings: Optimizing prompts reduces unnecessary back-and-forth and the expense of running inefficient inputs and outputs at scale.
Three practical techniques for businesses
Anthropic recommends three techniques that frequently unlock performance gains. They also stress that prompt engineering should begin with a clear task description and an experimental, iterative approach.
1) Step by step
Often the model gives more accurate answers if you instruct it to “think step by step” after the main instruction. For instance, when evaluating an automotive insurance claim, you can ask Claude to detail its decision process in <thinking> tags.
2) Few-shot prompting
Provide realistic, specific input–output examples, including difficult cases and edge scenarios, so the model understands the desired outputs. For tasks like removing personally identifiable information (PII), two examples showing different PII presentations help train the model on the correct redaction behavior.
3) Prompt chaining
Break complex tasks into multiple prompts corresponding to each step and include prior prompt–response pairs when needed. Prompt chaining lets you iteratively build context and guide Claude through multi-step processes — for example, collecting relevant tax codes first, then identifying pertinent sections, and finally answering user questions grounded in those excerpts.
Case study: improved accuracy and latency for a Fortune 500 customer
A Fortune 500 company sought an AI chat assistant to handle complex, specialized customer questions quickly and accurately. Prior market solutions produced wordy, stilted, and inconsistent output, and the client wanted lower latency without sacrificing accuracy.
They selected Claude Instant, but initial deployments failed to reach the desired accuracy. Anthropic dispatched a prompt engineer to collaborate with the company’s subject-matter experts to refine Claude’s responses. The prompt engineering team used several techniques:
- Requiring Claude to use a scratchpad that shows internal working (not visible to customers) to improve accuracy.
- Providing few-shot examples of preferred answers and training Claude to follow the company’s format and style.
- Directing Claude to rely on the data points and workflow recommended by subject-matter experts, such as the set of factors affecting a customer’s legal situation.
Combining prompt engineering best practices with SME guidance led to a 20% improvement in Claude’s accuracy, enabling faster time-to-market and lower costs for the customer.
Getting started
Anthropic notes there is no single best technique: each LLM behaves differently, and prompt strategies must be tested and iterated. For enterprise deployments, prompt engineering should be a collaborative effort between prompt engineers and subject-matter experts.
To begin optimizing Claude for a particular use case, Anthropic recommends consulting their prompt engineering guide or contacting their sales team for assistance.
Publication date
The article was published on February 29, 2024, and presents Anthropic’s recommendations and the described Fortune 500 case study.
