# Use an optimized prompt in the user's code

`client.prompts.get(task).text` is instruction text: no input placeholder, often literal JSON
braces. `.format(...)` raises `KeyError` and `.replace(...)` drops the input silently, which
ships a prompt that never sees the user's data.

Use `prompts.get(task).messages(**inputs)`. It renders the message shape the backend recorded
when it scored that version. The keyword names are that artifact's `input_variables`: `input`
for a plain task, `question` and `context` for RAG, or the user's own names if a template was
registered.

The score describes that shape sent to the model in `report.detail["student_model"]`, with
`response_format={"type": "json_object"}` for JSON metrics (`text={"format": ...}` on the
Responses API). Say so rather than implying the number transfers to any model.