Voice and format, not facts.

On the call, we would ask

• Is the problem how the tool writes, or what it knows?
• Do you have a few thousand good examples of the output you want?
• Who would re-run the tuning when the examples change?
• Would a good standing instruction get you most of the way?