Builds a condensed, numbered list of form questions for the LLM user prompt.
Strategy blocks are deliberately minimal — global extraction rules in the
system prompt already cover the common cases. Only type-specific constraints
that deviate from globals (option lists, scale bounds, group sub-fields) are
included per field. This reduces prompt token count by ~350-500 tokens on a
40-field form without sacrificing extraction accuracy.
required is deliberately NOT surfaced to the model — it conflicts with the
CONFIDENCE GATE rule ("UNCERTAIN evidence → OMIT, no exceptions"). Fields
with higher clinical priority are instead marked priority: high, which
encourages thoroughness without pressuring the model to fabricate.
Builds a condensed, numbered list of form questions for the LLM user prompt.
Strategy blocks are deliberately minimal — global extraction rules in the system prompt already cover the common cases. Only type-specific constraints that deviate from globals (option lists, scale bounds, group sub-fields) are included per field. This reduces prompt token count by ~350-500 tokens on a 40-field form without sacrificing extraction accuracy.
requiredis deliberately NOT surfaced to the model — it conflicts with the CONFIDENCE GATE rule ("UNCERTAIN evidence → OMIT, no exceptions"). Fields with higher clinical priority are instead markedpriority: high, which encourages thoroughness without pressuring the model to fabricate.