You are building an LLM feature that reads messy text such as emails, notes, or support tickets and pulls out a fixed set of fields for downstream analysis. The main issue is reliability: the model may miss fields, invent values, or return output that is hard to parse.
How would you design a prompt to reliably extract structured fields from text?
Prompt design for schema-constrained extractionHandling missing or ambiguous fields without guessingReducing hallucinated values with evidence requirementsMaking extraction quality measurable