Generative AI Prompt Engineering for Financial Text

Answer Strategy

Structure your answer using a clear framework: 1. **Decomposition**: Break the task into extraction (rating, outlook, rationale bullets) and synthesis (coherent summary). 2. **Prompt Strategy**: Use a system prompt to define the role as a 'neutral financial reporter'. Employ a two-step prompt chain: Step 1 extracts structured data; Step 2 synthesizes into prose. 3. **Validation**: Mention adding a final verification prompt that cross-checks the summary against the extracted data points for factual accuracy. Sample Answer: 'I'd build a two-stage pipeline. First, a structured extraction prompt with few-shot examples would parse the report for specific fields: rating, outlook, and a list of rationale points. Second, a synthesis prompt would take that JSON output and generate a concise, neutral paragraph. Finally, I'd implement a verification step where the model confirms every sentence in the summary is directly supported by the extracted data, using a self-critique loop.'

Answer Strategy

The interviewer is testing your debugging process, understanding of LLM failure modes, and ability to create robust systems. Focus on root cause analysis (e.g., ambiguous instructions, lack of source context, temperature too high) and systemic solutions (RAG, stricter output parsing, human-in-the-loop checks). Sample Answer: 'In a project generating earnings highlights, the model incorrectly stated a revenue figure from a prior period. The root cause was the prompt lacked a direct reference to the source data table. I fixed it by implementing Retrieval-Augmented Generation (RAG), where the prompt now includes the specific data table from the source document. Systemically, I mandated that all numerical claims in prompts must be grounded in attached source data, and I added an automated fact-checking layer that compares extracted numbers against a structured database.'