{"course":{"slug":"prompting","title":"Getting Real Work out of a Model","url":"https://addaly.com/learn/prompting","level":"beginner","track":"start"},"lesson":{"slug":"paste-the-source-not-a-summary","title":"Paste the source, not your account of it","url":"https://addaly.com/learn/prompting/paste-the-source-not-a-summary","position":11,"of":72,"minutes":8,"answer":"Describing your problem means deciding in advance what matters, so give the artefact and the surrounding facts and let your interpretation be a question rather than the input.","keyIdea":"Describing your problem means deciding in advance what matters, so give the artefact and the surrounding facts and let your interpretation be a question rather than the input."},"body":{"format":"markdown","text":"## Every summary is a decision about what matters\n\nWhen you describe your problem instead of showing it, you have already done the hardest part of the work, badly. You decided what was relevant. That decision is the thing you wanted help with.\n\n> My colleague sent a rude email about the project and I want to reply firmly but professionally.\n\nEvery important word there is your interpretation. *Rude* could be a sarcastic aside, a copied-in manager, or a flat sentence you have read four times in a bad mood. *The project* could be anything. The model now writes a reply to an email it has never seen, and it will be generic, because it is answering the only thing it was given: a category of situation.\n\nPaste the email. The reply changes completely, because the model can see the one sentence that actually stung and the three that were perfectly normal.\n\n## The three-layer rule\n\nFor any task involving a document, a message, a file or an error, there are three layers, and people habitually supply the wrong one.\n\n**Layer 1: the artefact.** The email. The contract clause. The traceback. The row of data. The photograph of the letter.\n**Layer 2: the surrounding facts.** Who sent it, when, what happened before, what changed recently.\n**Layer 3: your interpretation.** It seems rude. It looks like a bug in the loop. I think they are trying to get out of the contract.\n\nLayers 1 and 2 are what the model cannot get anywhere else. Layer 3 is worth including *as* an interpretation (\"I read this as hostile, am I right?\"), but never as a substitute for layer 1.\n\nThe single most common failure in this whole course is a prompt made entirely of layer 3.\n\n> Figure: The three layers of any document task. Layers, top to bottom: Layer 3: your interpretation ('it seems rude', 'I think they want out'. Include as a question to test.); Layer 2: the surrounding facts (who sent it, when, what happened before, what changed last week); Layer 1: the artefact (the email, the clause, the traceback, the row of data, pasted). Layers one and two are what the model cannot get anywhere else. Layer three belongs in the prompt as a question, never as the input. The commonest failure in the course is a prompt made entirely of layer three.\n\n## What \"the source\" means in different jobs\n\n**A bug.** The failing code, the full traceback, one line of the input data. Not \"it crashes on the sales file\". A traceback names the file, the line and the exception type; those three facts often solve it outright.\n\n**A document question.** The clause, plus enough either side that it is not quoted out of context. If the document is long, paste the section and say what the document is.\n\n**A data question.** The header row and five real rows, not a description of the columns. Column names carry information you have stopped noticing (`amt_inr`, `created_at_utc`, `status_2`) and the model will use them.\n\n**A writing task.** The draft, the brief you were given, and one piece of work you consider good. Do not describe the house style; show it.\n\n**A decision.** The actual options, with their actual numbers. Two quotes, both pasted, beat \"one is cheaper but slower\".\n\n## When you cannot paste\n\nSometimes you should not or cannot. A confidential contract, a patient record, a file too large for the window.\n\nThe honest options, in order of preference:\n\n1. **Redact and paste.** Replace names, account numbers and addresses with `[CLIENT]`, `[ACCOUNT]`, `[ADDRESS]`. The structure survives, the risk does not. This works for far more documents than people assume.\n2. **Paste the relevant part.** Section 9 and the two paragraphs around it.\n3. **Reconstruct a faithful example.** Write a fake invoice with the same shape as the real one. Genuinely useful for format and code problems, and useless for questions where the specific wording is the point.\n4. **Run it locally.** A small open model on your own machine, Ollama, LM Studio, `llama.cpp`, all free, never sends the text anywhere. The answers are weaker, and for many extraction and formatting jobs weaker is enough.\n\nWhat you should not do is describe the document in careful detail and hope. Six sentences of description are more work than a redacted paste and less use.\n\n## A test for whether you have pasted enough\n\nRead your prompt as if you had never seen the underlying material, and ask: **could I quote from the source to justify the answer?**\n\nIf the model can only produce an answer by relying on general knowledge of situations like yours, you have given it a category, not a case. General knowledge produces the advice you already knew, which brings us back to why the same question gets a better answer when it stops being a general question.\n\n## One caution\n\nPasting a source does not make an answer true. It changes what the answer is *about*. A model can still misread a clause, miss a negation, or over-read a polite sentence as a commitment. Later in the course you will ask it to quote the sentence it is relying on, which turns an interpretation you have to trust into a claim you can check in four seconds."},"check":{"question":"A team lead wants help replying to a supplier who \"is being difficult about the delivery date\". Which prompt is most likely to produce a usable reply?","options":[{"text":"A detailed description of the supplier relationship, the history of late deliveries, and what tone is needed.","correct":false,"tempting":"Rich in interpretation and missing the one thing only she has: the actual message she must answer."},{"text":"The same description plus three example replies she has sent to other suppliers.","correct":false,"tempting":"Examples fix tone, but the reply still has to respond to specific sentences the model has never seen."},{"text":"The supplier's message pasted in full, the date originally agreed, what changed since, and what outcome she needs.","correct":true},{"text":"A request for a general template for chasing late suppliers, which she will then adapt herself.","correct":false,"tempting":"Produces something reusable and generic, which is exactly the work she was trying to avoid doing by hand."}]},"prev":{"slug":"context-is-most-of-the-job","title":"Context is most of the job","url":"https://addaly.com/learn/prompting/context-is-most-of-the-job"},"next":{"slug":"marking-where-your-material-starts","title":"Marking where your material starts and stops","url":"https://addaly.com/learn/prompting/marking-where-your-material-starts"},"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) 'Paste the source, not your account of it', in Getting Real Work out of a Model. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/prompting/paste-the-source-not-a-summary (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":"Paste the source, not your account of it","courseTitle":"Getting Real Work out of a Model","module":{"index":2,"slug":"the-material-you-give-it","title":"The material you give it"},"answer":null,"keyIdea":"Describing your problem means deciding in advance what matters, so give the artefact and the surrounding facts and let your interpretation be a question rather than the input.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"type":"stack","title":"The three layers of any document task","layers":[{"label":"Layer 3: your interpretation","detail":"‘it seems rude’, ‘I think they want out’. Include as a question to test."},{"label":"Layer 2: the surrounding facts","detail":"who sent it, when, what happened before, what changed last week"},{"label":"Layer 1: the artefact","detail":"the email, the clause, the traceback, the row of data, pasted"}],"caption":"Layers one and two are what the model cannot get anywhere else. Layer three belongs in the prompt as a question, never as the input. The commonest failure in the course is a prompt made entirely of layer three."},"alt":"Figure: The three layers of any document task. Layers, top to bottom: Layer 3: your interpretation (‘it seems rude’, ‘I think they want out’. Include as a question to test.); Layer 2: the surrounding facts (who sent it, when, what happened before, what changed last week); Layer 1: the artefact (the email, the clause, the traceback, the row of data, pasted). Layers one and two are what the model cannot get anywhere else. Layer three belongs in the prompt as a question, never as the input. The commonest failure in the course is a prompt made entirely of layer three."}},"blocks":[{"t":"h2","id":"every-summary-is-a-decision-about-what-matters","runs":[{"text":"Every summary is a decision about what matters"}]},{"t":"p","runs":[{"text":"When you describe your problem instead of showing it, you have already done the hardest part of the work, badly. You decided what was relevant. That decision is the thing you wanted help with."}]},{"t":"quote","runs":[{"text":"My colleague sent a rude email about the project and I want to reply firmly but professionally."}]},{"t":"p","runs":[{"text":"Every important word there is your interpretation. "},{"text":"Rude","i":true},{"text":" could be a sarcastic aside, a copied-in manager, or a flat sentence you have read four times in a bad mood. "},{"text":"The project","i":true},{"text":" could be anything. The model now writes a reply to an email it has never seen, and it will be generic, because it is answering the only thing it was given: a category of situation."}]},{"t":"p","runs":[{"text":"Paste the email. The reply changes completely, because the model can see the one sentence that actually stung and the three that were perfectly normal."}]},{"t":"h2","id":"the-three-layer-rule","runs":[{"text":"The three-layer rule"}]},{"t":"p","runs":[{"text":"For any task involving a document, a message, a file or an error, there are three layers, and people habitually supply the wrong one."}]},{"t":"p","runs":[{"text":"Layer 1: the artefact.","b":true},{"text":" The email. The contract clause. The traceback. The row of data. The photograph of the letter. "},{"text":"Layer 2: the surrounding facts.","b":true},{"text":" Who sent it, when, what happened before, what changed recently. "},{"text":"Layer 3: your interpretation.","b":true},{"text":" It seems rude. It looks like a bug in the loop. I think they are trying to get out of the contract."}]},{"t":"p","runs":[{"text":"Layers 1 and 2 are what the model cannot get anywhere else. Layer 3 is worth including "},{"text":"as","i":true},{"text":" an interpretation (“I read this as hostile, am I right?”), but never as a substitute for layer 1."}]},{"t":"p","runs":[{"text":"The single most common failure in this whole course is a prompt made entirely of layer 3."}]},{"t":"h2","id":"what-the-source-means-in-different-jobs","runs":[{"text":"What “the source” means in different jobs"}]},{"t":"p","runs":[{"text":"A bug.","b":true},{"text":" The failing code, the full traceback, one line of the input data. Not “it crashes on the sales file”. A traceback names the file, the line and the exception type; those three facts often solve it outright."}]},{"t":"p","runs":[{"text":"A document question.","b":true},{"text":" The clause, plus enough either side that it is not quoted out of context. If the document is long, paste the section and say what the document is."}]},{"t":"p","runs":[{"text":"A data question.","b":true},{"text":" The header row and five real rows, not a description of the columns. Column names carry information you have stopped noticing ("},{"text":"amt_inr","code":true},{"text":", "},{"text":"created_at_utc","code":true},{"text":", "},{"text":"status_2","code":true},{"text":") and the model will use them."}]},{"t":"p","runs":[{"text":"A writing task.","b":true},{"text":" The draft, the brief you were given, and one piece of work you consider good. Do not describe the house style; show it."}]},{"t":"p","runs":[{"text":"A decision.","b":true},{"text":" The actual options, with their actual numbers. Two quotes, both pasted, beat “one is cheaper but slower”."}]},{"t":"h2","id":"when-you-cannot-paste","runs":[{"text":"When you cannot paste"}]},{"t":"p","runs":[{"text":"Sometimes you should not or cannot. A confidential contract, a patient record, a file too large for the window."}]},{"t":"p","runs":[{"text":"The honest options, in order of preference:"}]},{"t":"ol","start":1,"items":[[{"text":"Redact and paste.","b":true},{"text":" Replace names, account numbers and addresses with "},{"text":"[CLIENT]","code":true},{"text":", "},{"text":"[ACCOUNT]","code":true},{"text":", "},{"text":"[ADDRESS]","code":true},{"text":". The structure survives, the risk does not. This works for far more documents than people assume."}],[{"text":"Paste the relevant part.","b":true},{"text":" Section 9 and the two paragraphs around it."}],[{"text":"Reconstruct a faithful example.","b":true},{"text":" Write a fake invoice with the same shape as the real one. Genuinely useful for format and code problems, and useless for questions where the specific wording is the point."}],[{"text":"Run it locally.","b":true},{"text":" A small open model on your own machine, Ollama, LM Studio, "},{"text":"llama.cpp","code":true},{"text":", all free, never sends the text anywhere. The answers are weaker, and for many extraction and formatting jobs weaker is enough."}]]},{"t":"p","runs":[{"text":"What you should not do is describe the document in careful detail and hope. Six sentences of description are more work than a redacted paste and less use."}]},{"t":"h2","id":"a-test-for-whether-you-have-pasted-enough","runs":[{"text":"A test for whether you have pasted enough"}]},{"t":"p","runs":[{"text":"Read your prompt as if you had never seen the underlying material, and ask: "},{"text":"could I quote from the source to justify the answer?","b":true}]},{"t":"p","runs":[{"text":"If the model can only produce an answer by relying on general knowledge of situations like yours, you have given it a category, not a case. General knowledge produces the advice you already knew, which brings us back to why the same question gets a better answer when it stops being a general question."}]},{"t":"h2","id":"one-caution","runs":[{"text":"One caution"}]},{"t":"p","runs":[{"text":"Pasting a source does not make an answer true. It changes what the answer is "},{"text":"about","i":true},{"text":". A model can still misread a clause, miss a negation, or over-read a polite sentence as a commitment. Later in the course you will ask it to quote the sentence it is relying on, which turns an interpretation you have to trust into a claim you can check in four seconds."}]}],"check":{"question":[{"text":"A team lead wants help replying to a supplier who “is being difficult about the delivery date”. Which prompt is most likely to produce a usable reply?"}],"options":[{"text":[{"text":"A detailed description of the supplier relationship, the history of late deliveries, and what tone is needed."}],"correct":false,"tempting":[{"text":"Rich in interpretation and missing the one thing only she has: the actual message she must answer."}]},{"text":[{"text":"The same description plus three example replies she has sent to other suppliers."}],"correct":false,"tempting":[{"text":"Examples fix tone, but the reply still has to respond to specific sentences the model has never seen."}]},{"text":[{"text":"The supplier’s message pasted in full, the date originally agreed, what changed since, and what outcome she needs."}],"correct":true,"tempting":null},{"text":[{"text":"A request for a general template for chasing late suppliers, which she will then adapt herself."}],"correct":false,"tempting":[{"text":"Produces something reusable and generic, which is exactly the work she was trying to avoid doing by hand."}]}],"order":[3,0,2,1],"why":null},"prev":{"slug":"context-is-most-of-the-job","title":"Context is most of the job","minutes":7},"next":{"slug":"marking-where-your-material-starts","title":"Marking where your material starts and stops","minutes":7}}}