Summaries you can trust

A summary can leave out something important, add something that was never in the source, or sound sure about a shaky part. A few habits cut down all three.

Say who it is for and how long

Summarise this for someone deciding whether to read it, in five bullet points gives a different and more useful result than summarise this. Audience and length decide what stays.

Ask what was left out

Ask a second question: what in the original is important but did not make it into the summary? The model often names the points it dropped, which is quicker than comparing the two texts yourself.

Keep it inside the text

Add: use only information in the text below, and if the text does not say, write not stated. Without that sentence a model fills gaps from general knowledge, which reads well and may not be true of this document.

Keep quotes and numbers exact

Ask it to quote figures, names and dates directly and not to paraphrase them. Then spot-check three against the source. If they are right, the rest is more likely right. If one is off, read the source yourself.

For very long documents

Split the text, summarise the parts, then summarise the summaries. Details from the middle of a very long input are the ones most likely to be missed (see context-window).