{"course":{"slug":"prompting","title":"Getting Real Work out of a Model","url":"https://addaly.com/learn/prompting","level":"beginner","track":"start"},"lesson":{"slug":"reading-a-long-document","title":"Getting through a long document properly","url":"https://addaly.com/learn/prompting/reading-a-long-document","position":66,"of":72,"minutes":9,"answer":"Map the document and count its sections, locate the relevant ones, extract with quoted sources, verify two quotes, and summarise last: a summary produced first is a selection you cannot audit.","keyIdea":"Map the document and count its sections, locate the relevant ones, extract with quoted sources, verify two quotes, and summarise last: a summary produced first is a selection you cannot audit."},"body":{"format":"markdown","text":"## Summarising is the wrong first move\n\nThe instinct with a 60-page document is to ask for a summary. That gives you somebody else's selection of what matters, with no way to tell what was dropped, on a document you have not read.\n\nThere is a better sequence, and it takes about ten minutes.\n\n## 1. Map it\n\n> List every section heading in order, with its number and one line on what it covers. Do not summarise the content. Then give the total number of sections.\n\nYou get the shape of the document and a count. The count is your check against omission later, and the map tells you which three sections your question actually lives in.\n\nIf the map looks wrong (six headings for a document you know has thirty) you have caught a delivery problem before it cost you anything.\n\n## 2. Locate\n\n> Which sections deal with [your actual question]? List the section numbers and quote the first sentence of each. Do not analyse yet.\n\nCheap, hard to get wrong, and it turns a retrieval problem into a small reading problem.\n\n## 3. Extract, with sources\n\n> From sections 4, 9 and 17 only: list every obligation on us, with the clause number and the exact sentence. Include conditions and exceptions in the quoted sentence. If a section imposes no obligation, say so explicitly.\n\nThe \"say so explicitly\" clause is what stops a silent absence.\n\n## 4. Verify\n\nSearch two or three of the quotes in the original. Read the sentences either side of the one that matters most. Two minutes.\n\n## 5. Only now, summarise\n\n> Summarise the obligations above for [reader], who has five minutes and needs to know what we must do and by when.\n\nThe summary is now built on an enumerated, sourced, checked list rather than on a single pass over sixty pages. It is also much better, because the model is summarising a short structured input rather than compressing a long one.\n\n> Figure: Five passes over a sixty-page document. Steps: 1. Map (every heading in order, and a count of sections); 2. Locate (which sections carry the question; quote their first sentence); 3. Extract (every obligation with clause number and the exact sentence, qualifiers included); 4. Verify (search two quotes; read the sentences either side); 5. Summarise (only now, for the named reader). About ten minutes. The summary comes last, built on an enumerated, sourced and checked list rather than on one pass over material whose middle is used least reliably.\n\n## The questions worth asking of any long document\n\nBeyond your specific question, four sweeps catch most of what people miss:\n\n- \"List every deadline or date-triggered obligation, with its clause.\"\n- \"List every place where money changes hands, in either direction.\"\n- \"List every clause that lets either party end, suspend or change the arrangement.\"\n- \"List every place the document refers to another document, schedule or annex.\"\n\nThat last one catches the classic failure: the terms that matter are in Schedule 3, which nobody sent you.\n\n## What to do when it is a scan\n\nCheck first that it was read at all: \"quote the last sentence on page 12.\" If that fails, you have an OCR or delivery problem, not a comprehension problem, and no prompt will fix it. Extract the text yourself (`pdftotext` for a text-layer PDF, Tesseract for a scan, both free) or paste the pages that matter.\n\n## The honest scope of this\n\nYou now have a checked, sourced account of the parts you asked about. You do not have an assurance that nothing else matters, and you cannot get one from a model, because that judgement requires knowing what should be in a document of this kind.\n\nFor a tenancy agreement or a supplier contract, this workflow gets a careful non-specialist a long way. For anything where a missed clause is expensive (a shareholders' agreement, a construction contract, an employment settlement) it makes you a much better-prepared client for the person who reads it properly. That is a real gain, and it is a different gain from not needing them.\n\n## Why the order beats a better summary prompt\n\nIt is tempting to think the sequence could be collapsed if only the summarising instruction were sharper. It cannot, and the reason is structural rather than a limitation of phrasing.\n\nA single pass over sixty pages asks one process to locate, select, compress and explain at once, over material where the middle is used least reliably. The sequence separates those into four jobs, each over a smaller input, each producing an artefact you can check before the next one starts.\n\nThat is the same move as splitting a task, applied to reading. It is slower by about eight minutes and it is the difference between an account you can defend and one you are merely repeating."},"check":{"question":"Why is asking for the section map before anything else the highest-value first step?","options":[{"text":"It reduces the number of tokens needed for later questions.","correct":false,"tempting":"A minor efficiency, and not the reason the step changes outcomes."},{"text":"It improves the quality of the eventual summary by priming the model on the structure.","correct":false,"tempting":"Reverses cause and effect; the summary improves later because it is built on extracted material, not because of priming."},{"text":"It identifies which sections are most important in the document.","correct":false,"tempting":"Importance depends on your question; the map gives coverage, and you decide relevance."},{"text":"It reveals what actually reached the model and gives a count to check later answers against.","correct":true}]},"prev":{"slug":"the-message-you-do-not-want-to-write","title":"The message you have been putting off","url":"https://addaly.com/learn/prompting/the-message-you-do-not-want-to-write"},"next":{"slug":"comparing-options","title":"Deciding between options","url":"https://addaly.com/learn/prompting/comparing-options"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'Getting through a long document properly', in Getting Real Work out of a Model. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/prompting/reading-a-long-document (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Getting through a long document properly","courseTitle":"Getting Real Work out of a Model","module":{"index":8,"slug":"the-work-you-actually-have","title":"The work you actually have"},"answer":null,"keyIdea":"Map the document and count its sections, locate the relevant ones, extract with quoted sources, verify two quotes, and summarise last: a summary produced first is a selection you cannot audit.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"type":"flow","steps":[{"label":"Map","detail":"every heading in order, and a count of sections"},{"label":"Locate","detail":"which sections carry the question; quote their first sentence"},{"label":"Extract","detail":"every obligation with clause number and the exact sentence, qualifiers included"},{"label":"Verify","detail":"search two quotes; read the sentences either side"},{"label":"Summarise","detail":"only now, for the named reader"}],"title":"Five passes over a sixty-page document","caption":"About ten minutes. The summary comes last, built on an enumerated, sourced and checked list rather than on one pass over material whose middle is used least reliably."},"alt":"Figure: Five passes over a sixty-page document. Steps: 1. Map (every heading in order, and a count of sections); 2. Locate (which sections carry the question; quote their first sentence); 3. Extract (every obligation with clause number and the exact sentence, qualifiers included); 4. Verify (search two quotes; read the sentences either side); 5. Summarise (only now, for the named reader). About ten minutes. The summary comes last, built on an enumerated, sourced and checked list rather than on one pass over material whose middle is used least reliably."}},"blocks":[{"t":"h2","id":"summarising-is-the-wrong-first-move","runs":[{"text":"Summarising is the wrong first move"}]},{"t":"p","runs":[{"text":"The instinct with a 60-page document is to ask for a summary. That gives you somebody else’s selection of what matters, with no way to tell what was dropped, on a document you have not read."}]},{"t":"p","runs":[{"text":"There is a better sequence, and it takes about ten minutes."}]},{"t":"h2","id":"1-map-it","runs":[{"text":"1. Map it"}]},{"t":"quote","runs":[{"text":"List every section heading in order, with its number and one line on what it covers. Do not summarise the content. Then give the total number of sections."}]},{"t":"p","runs":[{"text":"You get the shape of the document and a count. The count is your check against omission later, and the map tells you which three sections your question actually lives in."}]},{"t":"p","runs":[{"text":"If the map looks wrong (six headings for a document you know has thirty) you have caught a delivery problem before it cost you anything."}]},{"t":"h2","id":"2-locate","runs":[{"text":"2. Locate"}]},{"t":"quote","runs":[{"text":"Which sections deal with [your actual question]? List the section numbers and quote the first sentence of each. Do not analyse yet."}]},{"t":"p","runs":[{"text":"Cheap, hard to get wrong, and it turns a retrieval problem into a small reading problem."}]},{"t":"h2","id":"3-extract-with-sources","runs":[{"text":"3. Extract, with sources"}]},{"t":"quote","runs":[{"text":"From sections 4, 9 and 17 only: list every obligation on us, with the clause number and the exact sentence. Include conditions and exceptions in the quoted sentence. If a section imposes no obligation, say so explicitly."}]},{"t":"p","runs":[{"text":"The “say so explicitly” clause is what stops a silent absence."}]},{"t":"h2","id":"4-verify","runs":[{"text":"4. Verify"}]},{"t":"p","runs":[{"text":"Search two or three of the quotes in the original. Read the sentences either side of the one that matters most. Two minutes."}]},{"t":"h2","id":"5-only-now-summarise","runs":[{"text":"5. Only now, summarise"}]},{"t":"quote","runs":[{"text":"Summarise the obligations above for [reader], who has five minutes and needs to know what we must do and by when."}]},{"t":"p","runs":[{"text":"The summary is now built on an enumerated, sourced, checked list rather than on a single pass over sixty pages. It is also much better, because the model is summarising a short structured input rather than compressing a long one."}]},{"t":"h2","id":"the-questions-worth-asking-of-any-long-document","runs":[{"text":"The questions worth asking of any long document"}]},{"t":"p","runs":[{"text":"Beyond your specific question, four sweeps catch most of what people miss:"}]},{"t":"ul","items":[[{"text":"“List every deadline or date-triggered obligation, with its clause.”"}],[{"text":"“List every place where money changes hands, in either direction.”"}],[{"text":"“List every clause that lets either party end, suspend or change the arrangement.”"}],[{"text":"“List every place the document refers to another document, schedule or annex.”"}]]},{"t":"p","runs":[{"text":"That last one catches the classic failure: the terms that matter are in Schedule 3, which nobody sent you."}]},{"t":"h2","id":"what-to-do-when-it-is-a-scan","runs":[{"text":"What to do when it is a scan"}]},{"t":"p","runs":[{"text":"Check first that it was read at all: “quote the last sentence on page 12.” If that fails, you have an OCR or delivery problem, not a comprehension problem, and no prompt will fix it. Extract the text yourself ("},{"text":"pdftotext","code":true},{"text":" for a text-layer PDF, Tesseract for a scan, both free) or paste the pages that matter."}]},{"t":"h2","id":"the-honest-scope-of-this","runs":[{"text":"The honest scope of this"}]},{"t":"p","runs":[{"text":"You now have a checked, sourced account of the parts you asked about. You do not have an assurance that nothing else matters, and you cannot get one from a model, because that judgement requires knowing what should be in a document of this kind."}]},{"t":"p","runs":[{"text":"For a tenancy agreement or a supplier contract, this workflow gets a careful non-specialist a long way. For anything where a missed clause is expensive (a shareholders’ agreement, a construction contract, an employment settlement) it makes you a much better-prepared client for the person who reads it properly. That is a real gain, and it is a different gain from not needing them."}]},{"t":"h2","id":"why-the-order-beats-a-better-summary-prompt","runs":[{"text":"Why the order beats a better summary prompt"}]},{"t":"p","runs":[{"text":"It is tempting to think the sequence could be collapsed if only the summarising instruction were sharper. It cannot, and the reason is structural rather than a limitation of phrasing."}]},{"t":"p","runs":[{"text":"A single pass over sixty pages asks one process to locate, select, compress and explain at once, over material where the middle is used least reliably. The sequence separates those into four jobs, each over a smaller input, each producing an artefact you can check before the next one starts."}]},{"t":"p","runs":[{"text":"That is the same move as splitting a task, applied to reading. It is slower by about eight minutes and it is the difference between an account you can defend and one you are merely repeating."}]}],"check":{"question":[{"text":"Why is asking for the section map before anything else the highest-value first step?"}],"options":[{"text":[{"text":"It reduces the number of tokens needed for later questions."}],"correct":false,"tempting":[{"text":"A minor efficiency, and not the reason the step changes outcomes."}]},{"text":[{"text":"It improves the quality of the eventual summary by priming the model on the structure."}],"correct":false,"tempting":[{"text":"Reverses cause and effect; the summary improves later because it is built on extracted material, not because of priming."}]},{"text":[{"text":"It identifies which sections are most important in the document."}],"correct":false,"tempting":[{"text":"Importance depends on your question; the map gives coverage, and you decide relevance."}]},{"text":[{"text":"It reveals what actually reached the model and gives a count to check later answers against."}],"correct":true,"tempting":null}],"order":[2,3,1,0],"why":null},"prev":{"slug":"the-message-you-do-not-want-to-write","title":"The message you have been putting off","minutes":9},"next":{"slug":"comparing-options","title":"Deciding between options","minutes":8}}}