{"course":{"slug":"prompting","title":"Getting Real Work out of a Model","url":"https://addaly.com/learn/prompting","level":"beginner","track":"start"},"lesson":{"slug":"specificity-and-its-limit","title":"How much detail is too much","url":"https://addaly.com/learn/prompting/specificity-and-its-limit","position":8,"of":72,"minutes":8,"answer":"Specificity pays until constraints start getting dropped or contradicting each other, so keep three to five checkable requirements and delete any sentence that would not change a correct answer.","keyIdea":"Specificity pays until constraints start getting dropped or contradicting each other, so keep three to five checkable requirements and delete any sentence that would not change a correct answer."},"body":{"format":"markdown","text":"## The advice has a ceiling\n\nEverything so far has pushed one direction: be more specific. That direction is right and it is where nearly all the gains are. But it does have a limit, and knowing where it sits will save you from the second-year mistake: the 600-word prompt with fourteen numbered rules that performs worse than the 120-word version.\n\nThree things go wrong when a prompt gets too dense.\n\n## 1. Constraints get silently dropped\n\nYou learned the mechanism two lessons ago: instruction-following is trained behaviour, not a checklist. In practice, past roughly five to eight hard constraints on a single output, models start satisfying most and ignoring some. Nothing announces the omission. The answer looks complete.\n\n> Figure: Constraints satisfied against constraints given. Line chart of Constraints actually satisfied against Hard constraints in the prompt: What you asked for (1, 1) (3, 3) (5, 5) (8, 8) (11, 11) (14, 14); What comes back (1, 1) (3, 3) (5, 4.8) (8, 6.5) (11, 8) (14, 9). Up to about five, everything holds. Past eight the gap opens and nothing announces it: eleven asked for, eight honoured, and the answer reads as complete.\n\nTest it and the pattern is easy to see: ask for a 100-word product description that mentions the warranty, avoids the word \"quality\", includes the SKU, uses British spelling, ends with a question, contains no adjectives before \"delivery\", and stays under three sentences. You will get a fluent paragraph missing two of those.\n\nWhat to do instead:\n\n- **Rank them.** \"These three are non-negotiable; the rest are preferences.\" Models honour a stated priority better than a flat list.\n- **Split the pass.** Draft first, then a second prompt that checks the draft against the list and fixes violations. Checking is an easier task than composing under constraint, and it is separately verifiable.\n- **Make them checkable.** \"Under 150 words\" you can verify at a glance. \"Engaging but not salesy\" you cannot, so it will drift and you will never quite know.\n\n## 2. Contradictions get resolved without telling you\n\n\"Comprehensive but under 200 words.\" \"Warm and professional but completely neutral in tone.\" \"Cover every clause but only mention the ones that matter.\"\n\nFaced with a contradiction, a model picks a reading and continues fluently. It rarely says *these two cannot both hold*. You are left with an answer that satisfies half of what you asked and reads as if it satisfied all of it.\n\nThe fix is a single line you can add to any complicated prompt:\n\n> Before answering, list any instructions above that conflict with each other or that you cannot satisfy together. Then answer, following the ones I marked as required.\n\nThis costs a few seconds and turns a silent compromise into a visible one.\n\n## 3. Detail that is not task-defining becomes noise\n\nThere is a real difference between two kinds of specificity.\n\n**Task-defining detail** changes what a correct answer is. The reader is a school governor with ten minutes. The budget is 400,000 pesos. The export changed last week. Remove any of these and the right answer changes.\n\n**Decorative detail** does not. The model's supposed twenty years of experience. Three sentences on why this project matters to you. A preamble explaining that you value clarity and precision. Remove them and the right answer is identical.\n\nDecorative detail is not neutral. It takes space in the context, and in a long prompt it dilutes the instructions that matter, because the model is spreading attention across everything present. A 400-word prompt where 250 words are throat-clearing is a worse prompt than the 150-word version.\n\nThe audit question for every sentence in your prompt: **would a correct answer change if I deleted this?** If not, delete it.\n\n## The exception worth knowing\n\nLong prompts are entirely appropriate when the length is *material* rather than instruction. A 4,000-word prompt that is 3,800 words of pasted contract and 200 words of task is well-proportioned. A 4,000-word prompt that is 3,800 words of rules about how to behave is usually a system that should have been split into stages.\n\nThe ratio to watch is instructions to material. Most weak prompts are heavy on instruction and starved of material; a few advanced ones tip the other way.\n\n## Where the balance actually sits\n\nFor everyday work: one to three sentences of task, as much genuine material as the job needs, three to five constraints that you can check, one line for the output shape, one line for the escape hatch. That will be somewhere between 80 and 300 words plus whatever you pasted, and it covers the overwhelming majority of real tasks.\n\nIf your prompt is longer than that and not mostly material, the question to ask is not \"what else should I add\" but \"which of these am I actually going to check\"."},"check":{"question":"A marketing assistant writes a 500-word prompt with twelve numbered rules for product copy. The outputs keep breaking two or three rules each time, but different ones. What is the most productive next move?","options":[{"text":"Add \"follow ALL twelve rules exactly, without exception\" at the top and bottom of the prompt.","correct":false,"tempting":"Emphasis does not create the checking step that is missing; the same rules will keep dropping out, just with more insistence in the prompt."},{"text":"Reduce the twelve rules to the three that matter and accept that the rest cannot be enforced.","correct":false,"tempting":"Closer, but it abandons nine requirements that a second checking pass could actually enforce."},{"text":"Provide five example outputs that follow all twelve rules, so the pattern transfers.","correct":false,"tempting":"Examples help with shape and tone but do not reliably transmit twelve simultaneous constraints, and they consume a great deal of context."},{"text":"Mark the three non-negotiable rules, then run a second prompt that checks the draft against all twelve and reports which are violated.","correct":true}]},"prev":{"slug":"the-folklore-audit","title":"Magic words, and what the studies actually found","url":"https://addaly.com/learn/prompting/the-folklore-audit"},"next":{"slug":"the-product-around-the-model","title":"Why the same prompt behaves differently in different apps","url":"https://addaly.com/learn/prompting/the-product-around-the-model"},"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) 'How much detail is too much', in Getting Real Work out of a Model. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/prompting/specificity-and-its-limit (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":"How much detail is too much","courseTitle":"Getting Real Work out of a Model","module":{"index":1,"slug":"what-a-prompt-is","title":"What a prompt actually is"},"answer":null,"keyIdea":"Specificity pays until constraints start getting dropped or contradicting each other, so keep three to five checkable requirements and delete any sentence that would not change a correct answer.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"type":"line","title":"Constraints satisfied against constraints given","series":[{"label":"What you asked for","points":[[1,1],[3,3],[5,5],[8,8],[11,11],[14,14]]},{"label":"What comes back","points":[[1,1],[3,3],[5,4.8],[8,6.5],[11,8],[14,9]]}],"xLabel":"Hard constraints in the prompt","yLabel":"Constraints actually satisfied","caption":"Up to about five, everything holds. Past eight the gap opens and nothing announces it: eleven asked for, eight honoured, and the answer reads as complete."},"alt":"Figure: Constraints satisfied against constraints given. Line chart of Constraints actually satisfied against Hard constraints in the prompt: What you asked for (1, 1) (3, 3) (5, 5) (8, 8) (11, 11) (14, 14); What comes back (1, 1) (3, 3) (5, 4.8) (8, 6.5) (11, 8) (14, 9). Up to about five, everything holds. Past eight the gap opens and nothing announces it: eleven asked for, eight honoured, and the answer reads as complete."}},"blocks":[{"t":"h2","id":"the-advice-has-a-ceiling","runs":[{"text":"The advice has a ceiling"}]},{"t":"p","runs":[{"text":"Everything so far has pushed one direction: be more specific. That direction is right and it is where nearly all the gains are. But it does have a limit, and knowing where it sits will save you from the second-year mistake: the 600-word prompt with fourteen numbered rules that performs worse than the 120-word version."}]},{"t":"p","runs":[{"text":"Three things go wrong when a prompt gets too dense."}]},{"t":"h2","id":"1-constraints-get-silently-dropped","runs":[{"text":"1. Constraints get silently dropped"}]},{"t":"p","runs":[{"text":"You learned the mechanism two lessons ago: instruction-following is trained behaviour, not a checklist. In practice, past roughly five to eight hard constraints on a single output, models start satisfying most and ignoring some. Nothing announces the omission. The answer looks complete."}]},{"t":"p","runs":[{"text":"Test it and the pattern is easy to see: ask for a 100-word product description that mentions the warranty, avoids the word “quality”, includes the SKU, uses British spelling, ends with a question, contains no adjectives before “delivery”, and stays under three sentences. You will get a fluent paragraph missing two of those."}]},{"t":"p","runs":[{"text":"What to do instead:"}]},{"t":"ul","items":[[{"text":"Rank them.","b":true},{"text":" “These three are non-negotiable; the rest are preferences.” Models honour a stated priority better than a flat list."}],[{"text":"Split the pass.","b":true},{"text":" Draft first, then a second prompt that checks the draft against the list and fixes violations. Checking is an easier task than composing under constraint, and it is separately verifiable."}],[{"text":"Make them checkable.","b":true},{"text":" “Under 150 words” you can verify at a glance. “Engaging but not salesy” you cannot, so it will drift and you will never quite know."}]]},{"t":"h2","id":"2-contradictions-get-resolved-without-telling-you","runs":[{"text":"2. Contradictions get resolved without telling you"}]},{"t":"p","runs":[{"text":"“Comprehensive but under 200 words.” “Warm and professional but completely neutral in tone.” “Cover every clause but only mention the ones that matter.”"}]},{"t":"p","runs":[{"text":"Faced with a contradiction, a model picks a reading and continues fluently. It rarely says "},{"text":"these two cannot both hold","i":true},{"text":". You are left with an answer that satisfies half of what you asked and reads as if it satisfied all of it."}]},{"t":"p","runs":[{"text":"The fix is a single line you can add to any complicated prompt:"}]},{"t":"quote","runs":[{"text":"Before answering, list any instructions above that conflict with each other or that you cannot satisfy together. Then answer, following the ones I marked as required."}]},{"t":"p","runs":[{"text":"This costs a few seconds and turns a silent compromise into a visible one."}]},{"t":"h2","id":"3-detail-that-is-not-task-defining-becomes-noise","runs":[{"text":"3. Detail that is not task-defining becomes noise"}]},{"t":"p","runs":[{"text":"There is a real difference between two kinds of specificity."}]},{"t":"p","runs":[{"text":"Task-defining detail","b":true},{"text":" changes what a correct answer is. The reader is a school governor with ten minutes. The budget is 400,000 pesos. The export changed last week. Remove any of these and the right answer changes."}]},{"t":"p","runs":[{"text":"Decorative detail","b":true},{"text":" does not. The model’s supposed twenty years of experience. Three sentences on why this project matters to you. A preamble explaining that you value clarity and precision. Remove them and the right answer is identical."}]},{"t":"p","runs":[{"text":"Decorative detail is not neutral. It takes space in the context, and in a long prompt it dilutes the instructions that matter, because the model is spreading attention across everything present. A 400-word prompt where 250 words are throat-clearing is a worse prompt than the 150-word version."}]},{"t":"p","runs":[{"text":"The audit question for every sentence in your prompt: "},{"text":"would a correct answer change if I deleted this?","b":true},{"text":" If not, delete it."}]},{"t":"h2","id":"the-exception-worth-knowing","runs":[{"text":"The exception worth knowing"}]},{"t":"p","runs":[{"text":"Long prompts are entirely appropriate when the length is "},{"text":"material","i":true},{"text":" rather than instruction. A 4,000-word prompt that is 3,800 words of pasted contract and 200 words of task is well-proportioned. A 4,000-word prompt that is 3,800 words of rules about how to behave is usually a system that should have been split into stages."}]},{"t":"p","runs":[{"text":"The ratio to watch is instructions to material. Most weak prompts are heavy on instruction and starved of material; a few advanced ones tip the other way."}]},{"t":"h2","id":"where-the-balance-actually-sits","runs":[{"text":"Where the balance actually sits"}]},{"t":"p","runs":[{"text":"For everyday work: one to three sentences of task, as much genuine material as the job needs, three to five constraints that you can check, one line for the output shape, one line for the escape hatch. That will be somewhere between 80 and 300 words plus whatever you pasted, and it covers the overwhelming majority of real tasks."}]},{"t":"p","runs":[{"text":"If your prompt is longer than that and not mostly material, the question to ask is not “what else should I add” but “which of these am I actually going to check”."}]}],"check":{"question":[{"text":"A marketing assistant writes a 500-word prompt with twelve numbered rules for product copy. The outputs keep breaking two or three rules each time, but different ones. What is the most productive next move?"}],"options":[{"text":[{"text":"Add “follow ALL twelve rules exactly, without exception” at the top and bottom of the prompt."}],"correct":false,"tempting":[{"text":"Emphasis does not create the checking step that is missing; the same rules will keep dropping out, just with more insistence in the prompt."}]},{"text":[{"text":"Reduce the twelve rules to the three that matter and accept that the rest cannot be enforced."}],"correct":false,"tempting":[{"text":"Closer, but it abandons nine requirements that a second checking pass could actually enforce."}]},{"text":[{"text":"Provide five example outputs that follow all twelve rules, so the pattern transfers."}],"correct":false,"tempting":[{"text":"Examples help with shape and tone but do not reliably transmit twelve simultaneous constraints, and they consume a great deal of context."}]},{"text":[{"text":"Mark the three non-negotiable rules, then run a second prompt that checks the draft against all twelve and reports which are violated."}],"correct":true,"tempting":null}],"order":[1,2,0,3],"why":null},"prev":{"slug":"the-folklore-audit","title":"Magic words, and what the studies actually found","minutes":9},"next":{"slug":"the-product-around-the-model","title":"Why the same prompt behaves differently in different apps","minutes":8}}}