These are structured so that the answer is checkable: you can look at the document and see whether a row is missing. That property matters more than any phrasing trick, because it converts "does this look right" into "is this right".
How to adapt them
- Replace every angle-bracket placeholder. A prompt left with
<goal>in it produces an answer about a goal the model invented. - Keep the "do not infer" line. It is the difference between an extract and a plausible-sounding reconstruction.
- Delimit pasted material. The triple quotes are not decoration; they mark where your instructions end and untrusted text begins. Text you paste can contain instructions of its own, and marking the boundary makes it far less likely they get followed.
When the answer is still wrong
If a table comes back with rows that are not in the document, that is a signal about the material; usually it is too long for one pass. Split it and run the extract per part, then merge your own extracted rows rather than the raw document.
What changes
Nothing here depends on a model version, which is deliberate. Re-test the structured ones after a major model update, since strictness about output format does shift between releases.