A prose-only brief for a housing ETL transform keeps flipping output formats. What is the most effective fix?
Select an answer to reveal the explanation.
Short Explanation
Prose that says "normalize nicely" is a weather vane in a storm. Nail the transform with concrete before-and-after examples and the format stops flipping.
Full Explanation
When a prose-only brief for a housing ETL transform keeps flipping output formats, the most effective fix is to attach concrete examples that show the expected transformation clearly. Format instability is a symptom of underspecified targets, and examples adjudicate format.
Concrete examples work because housing ETL outputs—column sets, date formats, address normalization, boolean encodings—are precise artifacts. A before-and-after pair (or a small set of pairs) tells Claude exactly which format is authoritative for the housing authority feed. Prose can describe goals; examples freeze the contract.
Adding more adjectives to the prose brief without samples fails because richer language still permits multiple formats. Restarting Claude Code with an empty context and the same prose fails by resetting memory while preserving the ambiguity that caused flips. Moving the brief into a user skill that still lacks examples fails by concept: relocating underspecified prose does not add the missing transformational contract.
Exam caveat: examples beat prose for format-sensitive ETL; skills help reuse, but only if they include the examples. Operational check: attach concrete housing ETL before/after examples, regenerate the transform, and verify the output format no longer flips across runs.