How to Ask an AI for Output You Can Check
A polished answer is hard to reuse. A schema with fields, types, and explicit gaps lets you validate what AI produced before passing it to another tool.
Prose is convenient; a schema is checkable
Asking “summarise this contract” may produce a useful answer, but it forces someone to find dates, copy amounts, and guess whether silence means no data existed or the AI omitted it. If the result goes into a spreadsheet, database, or review, ask for form before wording.
The capability to take away: turn an open task into a verifiable schema: fields, types, permitted values, and explicit treatment of missing data.
Define the container
JSON is a syntax for exchanging structured data, described in RFC 8259. It does not make an answer true; it makes its shape visible. To extract document obligations, define fields such as date, responsible_party, action, source, and extraction_confidence.
Then define rules: a date has a format; an amount includes currency; a category belongs to a list; an unknown field is null with an explanation. JSON Schema can express such restrictions.
Ask for gaps, not filler
The decisive instruction is: “Do not invent values. If the document lacks the datum, use null and explain what is missing.” This separates error from absence. Also request a reference to the source fragment for each material field.
A model can produce valid JSON with a false number. There are therefore two controls: validate that the structure follows the schema, then check important fields against the original document. NIST’s Generative AI Profile calls for measuring and managing risk; here that means not confusing “it imported” with “it is correct.”
A reusable template
State the task, schema, permitted values, what happens when data are missing, and required evidence. For example: “Extract obligations. Return only JSON. For each obligation: action, responsible party, ISO date or null, literal quote, and page. Do not infer; record ambiguities.”
The habit to keep
The more an answer will travel —to a person, spreadsheet, or automation— the more it should resemble a reviewable form and the less it should resemble elegant improvisation. First validate the form; then validate the truth of the data.
This article was produced with artificial intelligence under human editorial oversight.
