Summarize anything

The 40-page compliance circular has been sitting in Priya's inbox since Thursday, radiating quiet guilt. She will never read it properly — nobody reads the whole circular — but somewhere in it there may be three paragraphs that genuinely affect her fleet. This is the summarization problem, and most people solve it with AI badly, because they ask for the wrong thing: "summarize this document" produces a shorter version of everything, which is still about everything, which helps with nothing.

Summarize for the decision, not for coverage

The fix is to tell the assistant what you are going to do with the document. A summary is not a smaller document — it is an answer to the question you brought. Priya's version:

I run operations for a logistics company in Maharashtra — about 140
vehicles, no hazardous goods. Attached is a new compliance circular.

1. List every obligation that actually applies to a company like ours,
   with the section number for each.
2. Flag anything with a deadline, and the date.
3. Tell me what you'd check with a lawyer rather than assume.

Ignore everything about passenger transport and hazardous materials.

Three things make this work. The context ("140 vehicles, no hazardous goods") lets the model filter instead of compress — most of those 40 pages are now legitimately ignorable. The section numbers make every claim checkable in ten seconds, which is the difference between trusting a summary and being hostage to it. And the third question invites the model to mark its own uncertainty instead of papering over it — you will be surprised how often it takes the invitation.

The shapes a summary can take

"Summary" is a family of formats, and naming the one you want saves a revision:

• Action list — "what do I need to do, by when" — for circulars, policy updates, post-mortems.
• Decision brief — "what's being proposed, what are the options, what do they need from me" — for the 30-slide deck before the meeting.
• Key numbers — "pull every figure, with what it measures" — for reports you only need the data from.
• Both-sides brief — "the strongest case for and against" — for anything you have to form an opinion on.
• The one-liner — "one sentence: do I need to read this?" — the triage pass that decides whether to bother with the others.

For very long documents, ask in passes: the one-liner first, then a deep pass on only the sections that matter. It mirrors how a good analyst reads, and it keeps the model's attention where yours is.

Where the trap is

A summary is the model's reading, and the model can misread — it can miss a clause, blur two provisions together, or state a section number that's slightly off. So calibrate by stakes. For a conference agenda, take the summary and move on. For anything with legal, financial, or safety weight, use the summary as a map: it tells you which three sections to actually read, and you read them. Priya reads sections 12 and 31 herself — four pages instead of forty — and forwards the circular with a note that says exactly what changed for the fleet. That is the honest division of labour: the model does the reading you were never going to do, so your reading lands where it counts.

Do this today: take the longest unread document in your inbox and run the triage pass — "one sentence: do I need to read this?" — then, if yes, the decision-shaped summary with section references.