Prompt engineering
The craft of writing the instructions, examples and context given to a language model to obtain reliable, reproducible answers, which in companies becomes versioned, tested and governed documentation, like code.
Prompt engineering designates the design of the texts that steer a language model's behavior: description of the role and task, format constraints, examples of expected answers, context information, refusal rules, and breaking a complex task into steps. Small variations in wording can noticeably change the quality, tone or accuracy of answers, making this writing a skill in its own right. In professional use it is no longer an individual trick but system design. The instructions of an underwriting or claims assistant are assets that must be versioned, tested on representative case sets before each change, documented and approved, because they determine what the tool tells staff and sometimes customers. They must also be designed to resist hijacking attempts. For the insurer, governance of instructions joins governance of models: an untested instruction change can alter how thousands of files are analyzed, without any model being retrained, and so escape the usual controls.
Textbook case: a team changes the instruction of its policy analysis assistant so that it also flags communicable disease exclusions. Before going live, the new instruction is tested on 200 already annotated policies: it detects the new exclusions, but drops war exclusion detection from 96 to 88 percent, a flaw fixed before deployment.
Follow the research
One letter a week. First, the week's underwriting note, when one is published: a dated, argued decision, signed by Paul Bernardski. Then what the observatory published on cyber risk, political risk and insurance. No account, no tracking.
No spam, one-click unsubscribe, no data ever sold.