You have been told to check the AI's work, which is empty advice when the output is a bare list of dates with nothing to check them against. One line inside the prompt turns that list into something you can audit in two minutes with the contract open.
The same Austin agent publishes the seven workflows he runs on his own database, with the prompt for each. The fourth is the contract deadline one, and in Texas that means the option period, the financing deadline, and the closing date.
It opens plainly: "Find the executed contract for [ADDRESS] in my email. Read the whole document." Then it asks for the dates that matter, including "effective date, and how it is defined in this contract" and "option period end, including the exact day and time it expires".
Then the line that makes it usable:
"For each one, show me the date, the clause you took it from, and how you counted it."
"I want to check your math before anything goes on a calendar."
"If a date is ambiguous or a document is missing, say so instead of picking one."
"Then draft the calendar entries. Do not create them until I say go."
He is direct about why all of that is in there. The tool "is also confidently incorrect sometimes. A well formatted wrong closing date is worse than no closing date, because you will believe it." The rule sitting above every prompt in the piece is the same idea: "Never invent a transaction fact."
“For each one, show me the date, the clause you took it from, and how you counted it.”
An AI will hand you an option period end date with total confidence and nothing attached to it. Asking for the clause and the count converts an unverifiable answer into a two-minute check — you open the contract to that clause and confirm the arithmetic. That is the whole difference between something you can run on a live file and something you can only play with.
The second half is permission to say I do not know. A model produces an answer because producing answers is what it does, and one invented effective date moves every deadline that counts from it. Making an admission an allowed output is what stops a single guess from spreading through the entire file.