Open the folder where your team keeps its saved prompts and find two that do the same job. Then answer the harder question: how did you end up with both?<\/p>\n
What follows is a composite scenario, written for this article. A four-person customer support team uses AI to draft replies to refund requests. Two prompts in their shared folder produce that draft. Both work, both are in daily use, one is nine months old and one is five weeks old \u2014 and the team treats this as normal, because at every point along the way it was.<\/p>\n
The older prompt is titled Refund reply \u2014 standard<\/em>. It asks for a polite reply, references the returns window, and closes with an apology. The newer one is titled Refund draft v2 (Maya)<\/em>. It asks for the same reply, references the same window, and adds a line about expedited processing for orders above a certain value.<\/p>\n Neither title says what its prompt actually does, which is the first thing worth noticing: one is named for a category and one is named for a person, and a reader looking for the current version gets no help from either.<\/p>\n Read side by side, they are mostly identical. The small remaining difference is where everything important lives. Locating that difference requires more attention than most people give a prompt before choosing one and returning to the queue.<\/p>\n The path is short, and every step along it is defensible.<\/p>\n For six months there was one prompt. Someone wrote it, it worked, the team used it. This state looks stable, yet daily use is what eventually exposes the edge cases from which new versions grow.<\/p>\n In March the company introduced expedited processing for high-value orders. Maya, working a queue on a Tuesday, needed the draft to mention it. She had two options: edit the prompt three colleagues were using at that moment, or copy it and change her copy.<\/p>\n She copied it. This is the whole mechanism, and it deserves saying plainly: copying was the correct choice.<\/strong> Editing a shared prompt mid-shift changes what three other people get on their next request, with no warning and no easy way back. Copying is safe, local and reversible. Many prompt forks begin with exactly that trade: protect the shared version, or solve the immediate case locally.<\/p>\n Maya’s copy was better. It handled a case the original missed, so she kept using it, and when a colleague asked why her drafts mentioned expedited processing she sent him the link. Within a fortnight three of the four were on Maya’s copy.<\/p>\n The original was now superseded in practice and current in the shared folder. Those two facts sat beside each other for five months, because reconciling them sat outside everyone’s remit and both prompts were producing perfectly good replies.<\/p>\n Then Maya left. Her copy stayed in the folder under a title carrying her name, still in use by two people, still the better prompt, and now owned by no one. The fourth team member, who joined in June, has been using the original \u2014 because it is the one in the shared doc, and the shared doc is where a new person looks.<\/p>\n Every step in that sequence was sound. Maya solved a real problem the safest available way, her colleagues adopted an improvement, the new joiner followed the documented process. Judged one at a time, all four decisions hold up.<\/p>\n The cost appears only in the aggregate: a four-person team now sends customers two materially different refund replies depending on who picks up the ticket, and all four believe they are using the standard one. That is why the problem is difficult to assign to one decision or one person. It appears only when the separate choices are viewed as a portfolio.<\/p>\n Take two prompts that appear to do the same job and read them side by side. Identify every difference between them. If the effect of a difference remains unclear, run both prompts against the same few ordinary examples and compare what changes in the work. Then ask: does the difference reflect a real difference in the job?<\/em><\/p>\n That is the Prompt Pair Test. The answer sends you to one of three places, and all three are legitimate results.<\/p>\n Merge.<\/strong> The difference is an improvement one copy has and the other lacks. Maya’s expedited-processing rule belongs in the standard refund prompt as a conditional instruction for qualifying orders. One prompt should carry that rule forward and the other should leave active use.<\/p>\n Keep separate.<\/strong> The difference reflects two genuinely different jobs wearing similar titles. Had the older prompt handled refunds for a wholesale channel with its own terms, these would be two prompts badly named, and merging them would produce one that serves neither. The test earns its place precisely because it returns this answer sometimes<\/strong> \u2014 a consolidation drive that treats every pair as a duplicate destroys the distinction it failed to notice.<\/p>\n Retire.<\/strong> One prompt lost its job. If the returns window changed and only the newer text reflects it, the older one is obsolete, which differs from redundant in a way that matters. A merge carries something forward from both texts. A retirement leaves the survivor as it already was, and the retired prompt should leave the active folder, with its retirement and replacement recorded wherever the team keeps change history.<\/p>\n The support team ran it on their two refund prompts. The difference was one rule about expedited processing on qualifying orders, and it applied to the same job both prompts were doing. Merge. The team leaves with one approved refund prompt, a title describing the job rather than its author, and a recorded retirement for the older version.<\/p>\n The merge itself was quick. Updating the training notes that pointed to the retired prompt took considerably longer.<\/p>\n The following week the same team looked at Account cancellation reply<\/em> and Account cancellation \u2014 annual contract<\/em>. One colleague opened by calling the second one unnecessary duplication: same customer action, same folder, two prompts.<\/p>\n Its regular user disagreed, and her reason was operational. A monthly customer cancelling has a short notice period and a final bill that is simple to state. An annual customer cancelling mid-term has a notice period measured in weeks, a calculation against the unused portion of the term, and a question about whether cancellation takes effect now or at renewal. Her prompt asks for all of that. The general one asks for a polite confirmation.<\/p>\n Reading the two texts made the case arguable and left it unsettled, so they ran both against the same three cancellations: a monthly customer, an annual customer four months into a term, and an annual customer inside the renewal window. On the monthly cancellation the drafts were close. On the other two, the general prompt produced a confirmation that omitted the notice period and left the unused portion unmentioned \u2014 a reply the team would have had to follow up.<\/p>\n Keep separate. Both were renamed to say which customer they serve, and the colleague who had called it duplication wrote the new titles.<\/p>\n That outcome is worth as much as the merge.<\/strong> The person defending the second prompt held knowledge about a part of the job that the person proposing the cleanup had no reason to hold, and the disagreement is what brought it into the room. A consolidation decided on titles alone would have removed it.<\/p>\n You almost certainly have a candidate in mind \u2014 two prompts you have wondered about without ever making the time to look. Open both, read them side by side, find the difference, and decide whether that difference reflects a difference in the work.<\/p>\n One pair is enough to begin. The folder needs a first decision more than it needs an audit, and the second pair takes less time than the first. After a few pairs, the team begins to recognise which titles conceal likely overlap and which describe genuinely different work. Each later decision becomes faster because the folder<\/a> is becoming more legible, and a smaller portfolio is a different thing from a clearer one. If the same folder also has to survive a change of tool, that is a different exercise<\/a>.<\/p>\n","protected":false},"excerpt":{"rendered":" Open the folder where your team keeps its saved prompts and find two that do the same job. Then answer the harder question: how did you end up with both? What follows is a composite scenario, written for this article. A four-person customer support team uses AI to draft replies to refund requests. Two prompts in their shared folder produce that draft. Both work, both are in daily use, one is nine months old and one is five weeks old \u2014 and the team treats this as normal, because at every point along the way it was. The pair, as it stands today The older prompt is titled Refund reply \u2014 standard. It asks for a polite reply, references the returns window, and closes with an apology. The newer one is titled Refund draft v2 (Maya). It asks for the same reply, references the same window, and adds a line about expedited processing for orders above a certain value. Neither title says what its prompt actually does, which is the first thing worth noticing: one is named for a category and one is named for a person, and a reader looking for the current version gets no help from either. Read side by side, they are mostly identical. The small remaining difference is where everything important lives. Locating that difference requires more attention than most people give a prompt before choosing one and returning to the queue. How the second one appeared The path is short, and every step along it is defensible. In use For six months there was one prompt. Someone wrote it, it worked, the team used it. This state looks stable, yet daily use is what eventually exposes the edge cases from which new versions grow. Forked In March the company introduced expedited processing for high-value orders. Maya, working a queue on a Tuesday, needed the draft to mention it. She had two options: edit the prompt three colleagues were using at that moment, or copy it and change her copy. She copied it. This is the whole mechanism, and it deserves saying plainly: copying was the correct choice. Editing a shared prompt mid-shift changes what three other people get on their next request, with no warning and no easy way back. Copying is safe, local and reversible. Many prompt forks begin with exactly that trade: protect the shared version, or solve the immediate case locally. Superseded in practice, current in the folder Maya’s copy was better. It handled a case the original missed, so she kept using it, and when a colleague asked why her drafts mentioned expedited processing she sent him the link. Within a fortnight three of the four were on Maya’s copy. The original was now superseded in practice and current in the shared folder. Those two facts sat beside each other for five months, because reconciling them sat outside everyone’s remit and both prompts were producing perfectly good replies. Orphaned Then Maya left. Her copy stayed in the folder under a title carrying her name, still in use by two people, still the better prompt, and now owned by no one. The fourth team member, who joined in June, has been using the original \u2014 because it is the one in the shared doc, and the shared doc is where a new person looks. Where the cost actually sits Every step in that sequence was sound. Maya solved a real problem the safest available way, her colleagues adopted an improvement, the new joiner followed the documented process. Judged one at a time, all four decisions hold up. The cost appears only in the aggregate: a four-person team now sends customers two materially different refund replies depending on who picks up the ticket, and all four believe they are using the standard one. That is why the problem is difficult to assign to one decision or one person. It appears only when the separate choices are viewed as a portfolio. The Prompt Pair Test Take two prompts that appear to do the same job and read them side by side. Identify every difference between them. If the effect of a difference remains unclear, run both prompts against the same few ordinary examples and compare what changes in the work. Then ask: does the difference reflect a real difference in the job? That is the Prompt Pair Test. The answer sends you to one of three places, and all three are legitimate results. Merge. The difference is an improvement one copy has and the other lacks. Maya’s expedited-processing rule belongs in the standard refund prompt as a conditional instruction for qualifying orders. One prompt should carry that rule forward and the other should leave active use. Keep separate. The difference reflects two genuinely different jobs wearing similar titles. Had the older prompt handled refunds for a wholesale channel with its own terms, these would be two prompts badly named, and merging them would produce one that serves neither. The test earns its place precisely because it returns this answer sometimes \u2014 a consolidation drive that treats every pair as a duplicate destroys the distinction it failed to notice. Retire. One prompt lost its job. If the returns window changed and only the newer text reflects it, the older one is obsolete, which differs from redundant in a way that matters. A merge carries something forward from both texts. A retirement leaves the survivor as it already was, and the retired prompt should leave the active folder, with its retirement and replacement recorded wherever the team keeps change history. What the test produced here The support team ran it on their two refund prompts. The difference was one rule about expedited processing on qualifying orders, and it applied to the same job both prompts were doing. Merge. The team leaves with one approved refund prompt, a title describing the job rather than its author, and a recorded retirement for the older version. The merge itself was quick. Updating the training notes that pointed to the retired prompt took considerably longer. A second pair, correctly kept apart The following week the same team looked at Account cancellation reply and Account cancellation \u2014 annual contract. One colleague opened by calling the second one unnecessary duplication: same customer action, same folder, two prompts. Its regular user disagreed, and her reason was operational. A monthly customer cancelling has a short notice period and a final bill that is simple to state. An annual customer cancelling mid-term has a notice period measured in weeks, a calculation against the unused portion of the term, and a question about whether cancellation takes effect now or at renewal. Her prompt asks for all of that. The general one asks for a polite confirmation. Reading the two texts made the case arguable and left it unsettled, so they ran both against the same three cancellations: a monthly customer, an annual customer four months into a term, and an annual customer inside the renewal window. On the monthly cancellation the drafts were close. On the other two, the general prompt produced a confirmation that omitted the notice period and left the unused portion unmentioned \u2014 a reply the team would have had to follow up. Keep separate. Both were renamed to say which customer they serve, and the colleague who had called it duplication wrote the new titles. That outcome is worth as much as the merge. The person defending the second prompt held knowledge about a part of the job that the person proposing the cleanup had no reason to hold, and the disagreement is what brought it into the room. A consolidation decided on titles alone would have removed it. Start with the pair you already suspect You almost certainly have a candidate in mind \u2014 two prompts you have wondered about without ever making the time to look. Open both, read them side by side, find the difference, and decide whether that difference reflects a difference in the work. One pair is enough to begin. The folder needs a first decision more than it needs an audit, and the second pair takes less time than the first. After a few pairs, the team begins to recognise which titles conceal likely overlap and which describe genuinely different work. Each later decision becomes faster because the folder is becoming more legible, and a smaller portfolio is a different thing from a clearer one. If the same folder also has to survive a change of tool, that is a different exercise.<\/p>\n","protected":false},"author":113,"featured_media":34807,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"drb_seo_title":"Duplicate Prompts: Merge, Keep Separate, or Retire","drb_seo_desc":"Four people, two prompts, both working. Compare the pair and decide whether to merge them, keep them separate, or retire one.","footnotes":""},"categories":[1625],"tags":[],"class_list":["post-34534","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-in-practice"],"_links":{"self":[{"href":"https:\/\/dr-business.com\/en\/wp-json\/wp\/v2\/posts\/34534","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/dr-business.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/dr-business.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/dr-business.com\/en\/wp-json\/wp\/v2\/users\/113"}],"replies":[{"embeddable":true,"href":"https:\/\/dr-business.com\/en\/wp-json\/wp\/v2\/comments?post=34534"}],"version-history":[{"count":5,"href":"https:\/\/dr-business.com\/en\/wp-json\/wp\/v2\/posts\/34534\/revisions"}],"predecessor-version":[{"id":34793,"href":"https:\/\/dr-business.com\/en\/wp-json\/wp\/v2\/posts\/34534\/revisions\/34793"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/dr-business.com\/en\/wp-json\/wp\/v2\/media\/34807"}],"wp:attachment":[{"href":"https:\/\/dr-business.com\/en\/wp-json\/wp\/v2\/media?parent=34534"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/dr-business.com\/en\/wp-json\/wp\/v2\/categories?post=34534"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/dr-business.com\/en\/wp-json\/wp\/v2\/tags?post=34534"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}How the second one appeared<\/h2>\n
In use<\/h3>\n
Forked<\/h3>\n
Superseded in practice, current in the folder<\/h3>\n
Orphaned<\/h3>\n
Where the cost actually sits<\/h2>\n
The Prompt Pair Test<\/h2>\n
What the test produced here<\/h2>\n
A second pair, correctly kept apart<\/h2>\n
Start with the pair you already suspect<\/h2>\n