{"course":{"slug":"prompting","title":"Getting Real Work out of a Model","url":"https://addaly.com/learn/prompting","level":"beginner","track":"start"},"lesson":{"slug":"iterate-dont-reroll","title":"Edit the prompt, do not spin the wheel","url":"https://addaly.com/learn/prompting/iterate-dont-reroll","position":46,"of":72,"minutes":6,"answer":"Regenerating resamples the same prompt; naming what is wrong is the only thing that changes the pile.","keyIdea":"Regenerating resamples the same prompt; naming what is wrong is the only thing that changes the pile."},"body":{"format":"markdown","text":"## The regenerate button is a slot machine\n\nYou get a mediocre answer. You press regenerate. You get a differently mediocre answer. You press it four more times, pick the least bad one, and feel vaguely cheated.\n\nRegenerating draws another sample from roughly the same distribution. If the prompt is the problem, every draw comes from the same wrong pile. Occasionally you get a better one by luck, which is exactly what keeps people pulling the lever.\n\nThe alternative is to say what is wrong. Not \"make it better\" (that is regenerate with extra steps), but a specific diagnosis.\n\n## Vague correction versus real correction\n\n> Make it better.\n\n> Make it more engaging.\n\n> Try again, I don't like it.\n\nNone of these carry information. Compare:\n\n> Paragraph two assumes the reader knows what an SPV is. Define it in six words or cut the paragraph.\n\n> The tone is right but it is 40 percent too long. Cut the third and fifth bullets entirely and shorten the rest. Do not rewrite the opening, it is fine.\n\n> You invented a 20 percent discount. We do not offer discounts. Rewrite with the same structure, remove every mention of price.\n\nEach of these is a real edit. The last one is worth noticing: you are correcting a factual error, and if you do not name it, it comes back.\n\n## Two ways to iterate, and when to use each\n\n**Follow up in the same chat** when the output is mostly right and needs surgery. Cheap, fast, keeps all your context.\n\n**Edit the original prompt and start clean** when the conversation itself has gone wrong. This is the move people under-use.\n\nHere is why it matters. Everything in a conversation conditions what comes next. If turn three established a wrong assumption, turns four through thirty were built on top of it, and the wrong version is still sitting in the context competing with your correction. You are not overwriting it. You are arguing with it, in front of it.\n\nThe rule of thumb: **if you have corrected the same thing twice and it drifts back, stop correcting and start again** with the correction written into the first message.\n\n> In this file, the column \"net\" means revenue after refunds. It does not mean after tax. Here is the data.\n\nOne line at the top of a fresh chat beats five increasingly firm corrections at the bottom of a long one.\n\n## Watch for the sunk-cost chat\n\nA long conversation feels valuable. You have explained a lot. Starting over feels like throwing it away.\n\nBut the useful part of a long conversation is usually three or four sentences of context you can copy into a new one, and everything else is sediment. Copy the good part. Leave the rest.\n\n## Keep the ones that worked\n\nWhen a prompt finally produces what you wanted, save it. A plain text file, a note on your phone, anywhere.\n\nThis is the entirety of what \"prompt engineering\" means for most working people: a small file of prompts that produced good output, with the specifics swapped out next time. The four labelled examples from lesson four. The insurance table columns. The line that stops it inventing discounts.\n\nEach one took you a few minutes to work out and will save you those minutes every week after. Nobody else can write them for you, because they encode what your work actually needs."},"check":{"question":"Twenty turns into a conversation about a spreadsheet, the model keeps treating a column called \"net\" as revenue after tax, when in your file it means after refunds. You have corrected it twice and it has drifted back both times. What is the best move?","options":[{"text":"Correct it a third time, more forcefully and in capitals.","correct":false,"tempting":"When repetition fails, volume feels like the next lever, but the two earlier wrong usages are still sitting in the context, and shouting does not remove them."},{"text":"Rename the column in your file to something that cannot be misread, then re-paste it.","correct":false,"tempting":"Fixing the source data feels like the rigorous engineer's answer, but the model uses whatever definition you give it; the missing piece was a definition, not a better name."},{"text":"Start a fresh chat with the definition of \"net\" written into the first message, along with the few sentences of context that still matter.","correct":true},{"text":"Ask the model to summarise everything it has understood so far, then carry on from the summary.","correct":false,"tempting":"It sounds like good conversational hygiene, but the summary is built from the same contaminated context and tends to carry the wrong definition forward in a more compact form."}]},"prev":{"slug":"several-attempts","title":"Asking more than once, on purpose","url":"https://addaly.com/learn/prompting/several-attempts"},"next":{"slug":"diagnosing-a-bad-answer","title":"Four reasons an answer is bad","url":"https://addaly.com/learn/prompting/diagnosing-a-bad-answer"},"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) 'Edit the prompt, do not spin the wheel', in Getting Real Work out of a Model. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/prompting/iterate-dont-reroll (Accessed: 31 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Edit the prompt, do not spin the wheel","courseTitle":"Getting Real Work out of a Model","module":{"index":6,"slug":"from-lucky-to-reliable","title":"From lucky to reliable"},"answer":null,"keyIdea":"Regenerating resamples the same prompt; naming what is wrong is the only thing that changes the pile.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"the-regenerate-button-is-a-slot-machine","runs":[{"text":"The regenerate button is a slot machine"}]},{"t":"p","runs":[{"text":"You get a mediocre answer. You press regenerate. You get a differently mediocre answer. You press it four more times, pick the least bad one, and feel vaguely cheated."}]},{"t":"p","runs":[{"text":"Regenerating draws another sample from roughly the same distribution. If the prompt is the problem, every draw comes from the same wrong pile. Occasionally you get a better one by luck, which is exactly what keeps people pulling the lever."}]},{"t":"p","runs":[{"text":"The alternative is to say what is wrong. Not “make it better” (that is regenerate with extra steps), but a specific diagnosis."}]},{"t":"h2","id":"vague-correction-versus-real-correction","runs":[{"text":"Vague correction versus real correction"}]},{"t":"quote","runs":[{"text":"Make it better."}]},{"t":"quote","runs":[{"text":"Make it more engaging."}]},{"t":"quote","runs":[{"text":"Try again, I don’t like it."}]},{"t":"p","runs":[{"text":"None of these carry information. Compare:"}]},{"t":"quote","runs":[{"text":"Paragraph two assumes the reader knows what an SPV is. Define it in six words or cut the paragraph."}]},{"t":"quote","runs":[{"text":"The tone is right but it is 40 percent too long. Cut the third and fifth bullets entirely and shorten the rest. Do not rewrite the opening, it is fine."}]},{"t":"quote","runs":[{"text":"You invented a 20 percent discount. We do not offer discounts. Rewrite with the same structure, remove every mention of price."}]},{"t":"p","runs":[{"text":"Each of these is a real edit. The last one is worth noticing: you are correcting a factual error, and if you do not name it, it comes back."}]},{"t":"h2","id":"two-ways-to-iterate-and-when-to-use-each","runs":[{"text":"Two ways to iterate, and when to use each"}]},{"t":"p","runs":[{"text":"Follow up in the same chat","b":true},{"text":" when the output is mostly right and needs surgery. Cheap, fast, keeps all your context."}]},{"t":"p","runs":[{"text":"Edit the original prompt and start clean","b":true},{"text":" when the conversation itself has gone wrong. This is the move people under-use."}]},{"t":"p","runs":[{"text":"Here is why it matters. Everything in a conversation conditions what comes next. If turn three established a wrong assumption, turns four through thirty were built on top of it, and the wrong version is still sitting in the context competing with your correction. You are not overwriting it. You are arguing with it, in front of it."}]},{"t":"p","runs":[{"text":"The rule of thumb: "},{"text":"if you have corrected the same thing twice and it drifts back, stop correcting and start again","b":true},{"text":" with the correction written into the first message."}]},{"t":"quote","runs":[{"text":"In this file, the column “net” means revenue after refunds. It does not mean after tax. Here is the data."}]},{"t":"p","runs":[{"text":"One line at the top of a fresh chat beats five increasingly firm corrections at the bottom of a long one."}]},{"t":"h2","id":"watch-for-the-sunk-cost-chat","runs":[{"text":"Watch for the sunk-cost chat"}]},{"t":"p","runs":[{"text":"A long conversation feels valuable. You have explained a lot. Starting over feels like throwing it away."}]},{"t":"p","runs":[{"text":"But the useful part of a long conversation is usually three or four sentences of context you can copy into a new one, and everything else is sediment. Copy the good part. Leave the rest."}]},{"t":"h2","id":"keep-the-ones-that-worked","runs":[{"text":"Keep the ones that worked"}]},{"t":"p","runs":[{"text":"When a prompt finally produces what you wanted, save it. A plain text file, a note on your phone, anywhere."}]},{"t":"p","runs":[{"text":"This is the entirety of what “prompt engineering” means for most working people: a small file of prompts that produced good output, with the specifics swapped out next time. The four labelled examples from lesson four. The insurance table columns. The line that stops it inventing discounts."}]},{"t":"p","runs":[{"text":"Each one took you a few minutes to work out and will save you those minutes every week after. Nobody else can write them for you, because they encode what your work actually needs."}]}],"check":{"question":[{"text":"Twenty turns into a conversation about a spreadsheet, the model keeps treating a column called “net” as revenue after tax, when in your file it means after refunds. You have corrected it twice and it has drifted back both times. What is the best move?"}],"options":[{"text":[{"text":"Correct it a third time, more forcefully and in capitals."}],"correct":false,"tempting":[{"text":"When repetition fails, volume feels like the next lever, but the two earlier wrong usages are still sitting in the context, and shouting does not remove them."}]},{"text":[{"text":"Rename the column in your file to something that cannot be misread, then re-paste it."}],"correct":false,"tempting":[{"text":"Fixing the source data feels like the rigorous engineer’s answer, but the model uses whatever definition you give it; the missing piece was a definition, not a better name."}]},{"text":[{"text":"Start a fresh chat with the definition of “net” written into the first message, along with the few sentences of context that still matter."}],"correct":true,"tempting":null},{"text":[{"text":"Ask the model to summarise everything it has understood so far, then carry on from the summary."}],"correct":false,"tempting":[{"text":"It sounds like good conversational hygiene, but the summary is built from the same contaminated context and tends to carry the wrong definition forward in a more compact form."}]}],"order":[3,0,2,1],"why":null},"prev":{"slug":"several-attempts","title":"Asking more than once, on purpose","minutes":8},"next":{"slug":"diagnosing-a-bad-answer","title":"Four reasons an answer is bad","minutes":9}}}