{"course":{"slug":"ai-for-small-business","title":"AI for a Small Business","url":"https://addaly.com/learn/ai-for-small-business","level":"beginner","track":"work"},"lesson":{"slug":"fixing-a-bad-draft","title":"Repairing a draft instead of starting again","url":"https://addaly.com/learn/ai-for-small-business/fixing-a-bad-draft","position":14,"of":88,"minutes":8,"answer":"Repair a draft by naming the specific defect rather than asking for something better, stop after two failed turns.","keyIdea":"Repair a draft by naming the specific defect rather than asking for something better, stop after two failed turns, and promote any correction you make three times into the prompt itself so you never make it again."},"body":{"format":"markdown","text":"## Do not say \"try again\"\n\nThe first draft is wrong in some specific way. Most people type \"make it better\" or start a new chat. Both waste the useful thing you now have, which is a concrete example of the failure.\n\nName the defect precisely and ask for that one change. \"Too formal\" is vague; the model's idea of less formal may be contractions and an emoji. \"Remove the closing question and cut the first paragraph: start at the price\" is a specific instruction that lands the first time.\n\nA rough rule: if you can point at the sentence that is wrong, say what is wrong with that sentence. If you cannot, the problem is upstream in your facts or examples, and no amount of editing instructions will fix it.\n\n## The repairs that come up constantly\n\nFour patterns cover most business writing, and it is worth having the words ready:\n\n**It is too long.** Give a number. \"Cut to 60 words, keep the date and the price.\" Saying what to keep matters as much as the count, or it will drop the load-bearing detail and keep the pleasantries.\n\n**It sounds generated.** Usually three things: a closing rhetorical question, adjective stacking, and the smoothing away of the one odd specific detail. Say so directly. \"Delete the last line. Remove one adjective per sentence. Put the 40-minute wait back in.\"\n\n**It promised something.** \"Remove the refund offer. I did not agree to that. Rewrite the second paragraph so it commits to nothing beyond looking into it.\" Be blunt; this is the repair that protects money.\n\n**It missed the point.** The customer asked about Saturday delivery and the reply is about opening hours. Repair the input, not the output: the question was ambiguous, or the facts sheet has a gap. This one is a signal about your setup rather than about the draft.\n\n## Two turns, then stop\n\nIf the third attempt is still wrong, stop editing. Something structural is off: your request is unclear, the facts are missing, or the task is not one this tool does well.\n\nGoing round a fourth and fifth time feels like progress because each version differs, and it usually is not. Long repair threads also accumulate contradictory instructions: you asked for shorter, then for the deadline back, then for warmth, and the model is now trying to satisfy all three at once and doing none of them well.\n\nWhen you hit that point, start fresh with a better brief that incorporates what you learned. Ninety seconds, and it is usually right immediately.\n\n## Rewrite the prompt, not the draft\n\nThe move that separates people who get value from these tools from people who spend their day arguing with them.\n\nIf you find yourself making the same correction for the third time (it keeps using exclamation marks, it keeps inviting the customer to \"reach out\", it keeps inventing a delivery time) that correction belongs in the prompt. Add it as a permanent constraint, or add it to your standing instructions where it applies to everything.\n\nEach fix you promote from the draft to the prompt is a correction you never make again. Over a month this is the difference between a tool that saves an hour a week and one that costs you one.\n\n## When a fresh chat is the answer\n\nStart a new conversation when the thread has become contradictory, when you switch to an unrelated task, or when the answers start drifting oddly. Long threads carry everything said earlier, including your abandoned attempts and the model's wrong turns, all of which shape what comes next.\n\nCarry forward what is worth keeping. Ask for a two-line summary of the decisions made, paste it into the new chat with your facts sheet, and continue. The next lesson explains why long threads degrade in the first place, and why this handoff is cheaper than it looks.\n\n## Ask for three, not one\n\nWhen you know the draft is wrong but cannot say why, stop trying to describe it and ask for range instead: *give me three versions: one plainer, one shorter, one that leads with the price.*\n\nJudging three against each other is much easier than judging one in a vacuum, and you will often find the thing you wanted was in version two all along. It costs the same and takes the same time. This is also the fastest way to discover your own preference, which you can then write into the prompt as a permanent constraint.\n\n## Know when to stop and type\n\nSometimes the fastest repair is your own hands. If the draft is 80% right and the remaining 20% is a sentence you can write in fifteen seconds, write it. People who are new to these tools often spend three minutes instructing a machine to make a change they could have made in ten seconds, because it feels like the tool ought to do it.\n\nThe tool is for the part that is long and tedious. The last sentence is neither. A good working rhythm is: one repair instruction, then finish it yourself."},"check":{"question":"An owner has now told the model four times in one thread to stop signing off with \"Feel free to reach out!\". Each time it complies, and the phrase returns two drafts later. What is the right response?","options":[{"text":"Start a fresh chat, since the thread has accumulated contradictory instructions and the sign-off keeps returning because earlier drafts containing it are still sitting in context.","correct":false,"tempting":"Clears the symptom for one session while losing the fix, so the same phrase returns in the next fresh chat and the correction has to be made a fifth time."},{"text":"Move the instruction into the standing prompt or facts sheet, so it applies to every draft rather than being re-issued each time it reappears.","correct":true},{"text":"Ask the model to explain why it keeps reintroducing the phrase, so the underlying cause can be addressed.","correct":false,"tempting":"Treats the model as a reliable narrator of its own behaviour, when any explanation it offers is generated text rather than an inspection of what happened."},{"text":"Switch to a different model, since this one is evidently ignoring instructions.","correct":false,"tempting":"Changes vendor to solve a prompt-persistence problem that every model has, because a one-off instruction earlier in a long thread carries less weight than a standing one."}]},"prev":{"slug":"making-it-admit-ignorance","title":"Making the gaps visible","url":"https://addaly.com/learn/ai-for-small-business/making-it-admit-ignorance"},"next":{"slug":"context-and-forgetting","title":"Why long conversations go strange","url":"https://addaly.com/learn/ai-for-small-business/context-and-forgetting"},"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) 'Repairing a draft instead of starting again', in AI for a Small Business. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-for-small-business/fixing-a-bad-draft (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":"Repairing a draft instead of starting again","courseTitle":"AI for a Small Business","module":{"index":2,"slug":"asking-properly","title":"Asking properly: prompting for business work"},"answer":null,"keyIdea":"Repair a draft by naming the specific defect rather than asking for something better, stop after two failed turns, and promote any correction you make three times into the prompt itself so you never make it again.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"do-not-say-try-again","runs":[{"text":"Do not say “try again”"}]},{"t":"p","runs":[{"text":"The first draft is wrong in some specific way. Most people type “make it better” or start a new chat. Both waste the useful thing you now have, which is a concrete example of the failure."}]},{"t":"p","runs":[{"text":"Name the defect precisely and ask for that one change. “Too formal” is vague; the model’s idea of less formal may be contractions and an emoji. “Remove the closing question and cut the first paragraph: start at the price” is a specific instruction that lands the first time."}]},{"t":"p","runs":[{"text":"A rough rule: if you can point at the sentence that is wrong, say what is wrong with that sentence. If you cannot, the problem is upstream in your facts or examples, and no amount of editing instructions will fix it."}]},{"t":"h2","id":"the-repairs-that-come-up-constantly","runs":[{"text":"The repairs that come up constantly"}]},{"t":"p","runs":[{"text":"Four patterns cover most business writing, and it is worth having the words ready:"}]},{"t":"p","runs":[{"text":"It is too long.","b":true},{"text":" Give a number. “Cut to 60 words, keep the date and the price.” Saying what to keep matters as much as the count, or it will drop the load-bearing detail and keep the pleasantries."}]},{"t":"p","runs":[{"text":"It sounds generated.","b":true},{"text":" Usually three things: a closing rhetorical question, adjective stacking, and the smoothing away of the one odd specific detail. Say so directly. “Delete the last line. Remove one adjective per sentence. Put the 40-minute wait back in.”"}]},{"t":"p","runs":[{"text":"It promised something.","b":true},{"text":" “Remove the refund offer. I did not agree to that. Rewrite the second paragraph so it commits to nothing beyond looking into it.” Be blunt; this is the repair that protects money."}]},{"t":"p","runs":[{"text":"It missed the point.","b":true},{"text":" The customer asked about Saturday delivery and the reply is about opening hours. Repair the input, not the output: the question was ambiguous, or the facts sheet has a gap. This one is a signal about your setup rather than about the draft."}]},{"t":"h2","id":"two-turns-then-stop","runs":[{"text":"Two turns, then stop"}]},{"t":"p","runs":[{"text":"If the third attempt is still wrong, stop editing. Something structural is off: your request is unclear, the facts are missing, or the task is not one this tool does well."}]},{"t":"p","runs":[{"text":"Going round a fourth and fifth time feels like progress because each version differs, and it usually is not. Long repair threads also accumulate contradictory instructions: you asked for shorter, then for the deadline back, then for warmth, and the model is now trying to satisfy all three at once and doing none of them well."}]},{"t":"p","runs":[{"text":"When you hit that point, start fresh with a better brief that incorporates what you learned. Ninety seconds, and it is usually right immediately."}]},{"t":"h2","id":"rewrite-the-prompt-not-the-draft","runs":[{"text":"Rewrite the prompt, not the draft"}]},{"t":"p","runs":[{"text":"The move that separates people who get value from these tools from people who spend their day arguing with them."}]},{"t":"p","runs":[{"text":"If you find yourself making the same correction for the third time (it keeps using exclamation marks, it keeps inviting the customer to “reach out”, it keeps inventing a delivery time) that correction belongs in the prompt. Add it as a permanent constraint, or add it to your standing instructions where it applies to everything."}]},{"t":"p","runs":[{"text":"Each fix you promote from the draft to the prompt is a correction you never make again. Over a month this is the difference between a tool that saves an hour a week and one that costs you one."}]},{"t":"h2","id":"when-a-fresh-chat-is-the-answer","runs":[{"text":"When a fresh chat is the answer"}]},{"t":"p","runs":[{"text":"Start a new conversation when the thread has become contradictory, when you switch to an unrelated task, or when the answers start drifting oddly. Long threads carry everything said earlier, including your abandoned attempts and the model’s wrong turns, all of which shape what comes next."}]},{"t":"p","runs":[{"text":"Carry forward what is worth keeping. Ask for a two-line summary of the decisions made, paste it into the new chat with your facts sheet, and continue. The next lesson explains why long threads degrade in the first place, and why this handoff is cheaper than it looks."}]},{"t":"h2","id":"ask-for-three-not-one","runs":[{"text":"Ask for three, not one"}]},{"t":"p","runs":[{"text":"When you know the draft is wrong but cannot say why, stop trying to describe it and ask for range instead: "},{"text":"give me three versions: one plainer, one shorter, one that leads with the price.","i":true}]},{"t":"p","runs":[{"text":"Judging three against each other is much easier than judging one in a vacuum, and you will often find the thing you wanted was in version two all along. It costs the same and takes the same time. This is also the fastest way to discover your own preference, which you can then write into the prompt as a permanent constraint."}]},{"t":"h2","id":"know-when-to-stop-and-type","runs":[{"text":"Know when to stop and type"}]},{"t":"p","runs":[{"text":"Sometimes the fastest repair is your own hands. If the draft is 80% right and the remaining 20% is a sentence you can write in fifteen seconds, write it. People who are new to these tools often spend three minutes instructing a machine to make a change they could have made in ten seconds, because it feels like the tool ought to do it."}]},{"t":"p","runs":[{"text":"The tool is for the part that is long and tedious. The last sentence is neither. A good working rhythm is: one repair instruction, then finish it yourself."}]}],"check":{"question":[{"text":"An owner has now told the model four times in one thread to stop signing off with “Feel free to reach out!”. Each time it complies, and the phrase returns two drafts later. What is the right response?"}],"options":[{"text":[{"text":"Start a fresh chat, since the thread has accumulated contradictory instructions and the sign-off keeps returning because earlier drafts containing it are still sitting in context."}],"correct":false,"tempting":[{"text":"Clears the symptom for one session while losing the fix, so the same phrase returns in the next fresh chat and the correction has to be made a fifth time."}]},{"text":[{"text":"Move the instruction into the standing prompt or facts sheet, so it applies to every draft rather than being re-issued each time it reappears."}],"correct":true,"tempting":null},{"text":[{"text":"Ask the model to explain why it keeps reintroducing the phrase, so the underlying cause can be addressed."}],"correct":false,"tempting":[{"text":"Treats the model as a reliable narrator of its own behaviour, when any explanation it offers is generated text rather than an inspection of what happened."}]},{"text":[{"text":"Switch to a different model, since this one is evidently ignoring instructions."}],"correct":false,"tempting":[{"text":"Changes vendor to solve a prompt-persistence problem that every model has, because a one-off instruction earlier in a long thread carries less weight than a standing one."}]}],"order":[3,2,0,1],"why":null},"prev":{"slug":"making-it-admit-ignorance","title":"Making the gaps visible","minutes":9},"next":{"slug":"context-and-forgetting","title":"Why long conversations go strange","minutes":8}}}