Playbook, documents

Turn a 40-page document into actions you trust

The three sentences that matter are on page 31. Our extraction routine finds them and proves where they came from.

The Curio Team · 2 min readWritten with AI from the sources below and checked against them before publishing.

Nobody reads the 40-page report. People read the first two pages, skim the headers, and hope. The decisions hide deeper: a deadline in a footnote, a cost assumption on page 31, an owner nobody confirmed. We stopped skimming and started extracting. Working with AI on long documents is one rule wrapped in a routine: the model can summarize anything, so never let it summarize. Make it quote.

Placement and grounding do real work here. Anthropic's long-context guidance says to put the document at the top of the prompt and ask the model to pull relevant quotes before answering, because answers grounded in quoted text drift less than answers from its memory of the text.¹

Structure does the rest. OpenAI's text generation docs push the same practical point from their side: specify the output format exactly, because a loose ask returns a loose answer. So we require fixed fields every time: decision needed, deadline, owner, dollar amount, quote.²

The routine, start to finish:

1. Get the map first. Paste the document and ask for a section-by-section outline of what each part claims. Two minutes here tells you whether the model actually ingested page 31 or quietly truncated.
2. Extract into fixed fields. Every decision needed, deadline, owner, and dollar amount, each with the exact sentence it came from. Same fields for every document, so your eye learns where to look.
3. Enforce quote-or-cut. If the model cannot produce the supporting sentence, the row goes. This one rule converts a fluent summarizer into an honest extractor.
4. Interrogate the gaps. Ask what a careful reader would expect this document to address that it never does. Missing risks and unowned actions surface here, and they are usually the real story.
5. Spot-check three quotes against the actual pages before you act. Thirty seconds each. If one fails, distrust the batch, split the document in half, and rerun.

No quote, no row.

A transcript gets the same treatment as a report. Meeting notes are just a long document where the owners talk back, and the fields shrink to task, owner, and date. The one difference is volume: transcripts repeat themselves, so the map step matters more, not less.

Length is the honest limit. A very long document can exceed what the model reads faithfully in one pass, and the failure is silent: it paraphrases the middle instead of quoting it. The section map catches this early. If the outline skips a chapter you know exists, split the file in half and run each part separately.¹

The payoff is trust you can defend. When the action list carries page-level quotes, the Thursday meeting stops arguing about whether the summary is right and starts deciding. That is the whole point of putting a model to work on your documents: it does the reading, we do the deciding, and the quotes are the handshake between us.