Why a filled-in brief is the wrong thing to automate
Ask a language model to write a pre-meeting brief about a named household and it will produce one. That is the problem, not the reassurance. A model has no way to say 'I don't know' by default; asked for a paragraph about a specific person, it fills every gap with the most statistically plausible detail, and a plausible detail about a real person that did not happen is not a rounding error, it is a false statement about someone you are about to sit across from.
The only fix that survives contact with a real household is structural, not a politer prompt. This tool's schema does not have a field for 'what the household is like'. It has fields for the ordered steps of a method, for how you would corroborate a record, and for what to leave blank. There is no shape in the output for the model to hide a guess inside.