Key idea: A model can return the right information in a shape your code can't read. Enforce a schema so every reply parses, and describe the tricky fields so the values come back right.
Could the app use it?
Run a setting to add it here
- Run Plain words. The model got the information, so why couldn't the app read a single reply?
- Run Schema in prompt on Fast, then on Smart. Why does spelling out the shape work for one model and not the other?
- Run Enforced on Smart. Every reply parses, so what went wrong in the ones marked wrong?
- Run Enforced + descriptions. What did one line per field fix that the schema alone couldn't?