WEBVTT 1 00:00:00.225 --> 00:00:06.087 You're listening to a new approach to prompt writing that flips the structure by placing constraints at the top. 2 00:00:06.087 --> 00:00:09.225 Written by Maaren Holloway on Promptway. 3 00:00:09.760 --> 00:00:12.585 I'll save you the forty minutes I wasted on this. 4 00:00:12.860 --> 00:00:13.685 For about a year I 5 00:00:13.685 --> 00:00:15.997 wrote prompts the way every guide tells you to. 6 00:00:15.997 --> 00:00:20.085 Role at the top, context next, task in the middle, format and constraints at the bottom. 7 00:00:20.085 --> 00:00:21.085 It read clean. 8 00:00:21.085 --> 00:00:23.285 It also produced drafts I had to rewrite. 9 00:00:23.285 --> 00:00:27.697 Then I flipped the order, and the same model on the same input started giving me work I could ship. 10 00:00:27.697 --> 00:00:28.422 Same words, 11 00:00:28.422 --> 00:00:30.760 different position, different output. 12 00:00:31.035 --> 00:00:32.185 The thesis is simple. 13 00:00:32.185 --> 00:00:34.810 Constraints belong at the top of the prompt, not the bottom. 14 00:00:34.810 --> 00:00:37.422 The model is paying the most attention to what it reads first. 15 00:00:37.422 --> 00:00:37.522 If 16 00:00:37.522 --> 00:00:37.610 the 17 00:00:37.610 --> 00:00:37.822 first 18 00:00:37.822 --> 00:00:37.947 thing 19 00:00:37.947 --> 00:00:38.035 it 20 00:00:38.035 --> 00:00:38.335 reads 21 00:00:38.335 --> 00:00:38.472 is 22 00:00:38.472 --> 00:00:38.610 "you 23 00:00:38.610 --> 00:00:38.697 are 24 00:00:38.697 --> 00:00:38.760 a 25 00:00:38.760 --> 00:00:39.085 senior 26 00:00:39.085 --> 00:00:39.510 content 27 00:00:39.510 --> 00:00:40.310 strategist," 28 00:00:40.310 --> 00:00:40.397 it 29 00:00:40.397 --> 00:00:40.622 spends 30 00:00:40.622 --> 00:00:40.685 a 31 00:00:40.685 --> 00:00:41.260 paragraph 32 00:00:41.260 --> 00:00:41.447 being 33 00:00:41.447 --> 00:00:41.510 a 34 00:00:41.510 --> 00:00:41.835 senior 35 00:00:41.835 --> 00:00:42.260 content 36 00:00:42.260 --> 00:00:42.860 strategist 37 00:00:42.860 --> 00:00:43.122 before 38 00:00:43.122 --> 00:00:47.185 it considers that you needed a 90-word answer in plain English with no bullet points. 39 00:00:47.185 --> 00:00:50.722 By the time it gets to your rules, it has already committed to a shape. 40 00:00:51.010 --> 00:00:57.185 Then it negotiates with itself for the rest of the response, and you can feel the negotiation in the draft. 41 00:00:57.460 --> 00:00:59.610 This is not a personality quirk of Claude. 42 00:00:59.610 --> 00:01:09.160 There is now a small pile of research, including Anthropic's own guidance, pointing the same direction: telling the model what NOT to do, and putting the hardest constraints up front, does more work than any other single change you can make to a prompt. 43 00:01:09.160 --> 00:01:12.635 I have tested this on roughly two hundred client briefs over the last six months. 44 00:01:12.635 --> 00:01:15.860 The ordering change alone cut my rewrite time by something like a third. 45 00:01:15.860 --> 00:01:17.797 I did not measure it cleanly. 46 00:01:18.060 --> 00:01:20.085 I am not going to pretend I did. 47 00:01:20.360 --> 00:01:25.885 What the bad version looks like Here is a sanitized version of the prompt I used to send. 48 00:01:25.885 --> 00:01:31.185 I was writing a homepage hero brief for a Series A company in the supply-chain space. 49 00:01:31.185 --> 00:01:33.385 The client wanted something specific. 50 00:01:33.385 --> 00:01:35.610 They got something generic. 51 00:01:35.910 --> 00:01:41.297 What I got back was three headlines that all sounded like Stripe's third-most-confident competitor. 52 00:01:41.297 --> 00:01:50.910 Two of them used the word "seamless." One of them said "transform." Both were on the banned list, in the same prompt, three paragraphs above the output. 53 00:01:51.235 --> 00:01:53.210 The model was not being stupid. 54 00:01:53.210 --> 00:02:00.760 It was responding to the prompt I actually wrote, which spent its opening describing a persona that writes that kind of copy for a living. 55 00:02:00.760 --> 00:02:05.135 The constraints arrived after the model had already chosen a register. 56 00:02:05.410 --> 00:02:09.422 What the working version looks like Here is the same prompt restructured. 57 00:02:09.422 --> 00:02:10.235 Same facts. 58 00:02:10.235 --> 00:02:11.585 Different order. 59 00:02:11.910 --> 00:02:13.372 Three headlines came back. 60 00:02:13.372 --> 00:02:15.072 None of them used a banned word. 61 00:02:15.072 --> 00:02:20.960 Two were genuinely different in angle (one led on speed of integration, one led on the cost of a bad pick). 62 00:02:20.960 --> 00:02:23.060 The third was weaker, which is honest. 63 00:02:23.060 --> 00:02:25.335 The client picked one with a one-word edit. 64 00:02:25.335 --> 00:02:28.260 I sent the invoice the same afternoon. 65 00:02:28.560 --> 00:02:30.135 The change is not magic. 66 00:02:30.135 --> 00:02:35.885 The model treats the top of your prompt as the part it has to honor; the bottom is the part it negotiates with. 67 00:02:35.885 --> 00:02:41.647 If the part you cannot afford to lose is at the bottom, you are betting that the model will negotiate in your favor. 68 00:02:41.647 --> 00:02:43.010 Sometimes it will. 69 00:02:43.010 --> 00:02:45.710 On a tight brief, you do not want to bet. 70 00:02:45.985 --> 00:02:53.760 Why the order works Two things are happening, and they are worth understanding so you can adapt the pattern instead of memorizing it. 71 00:02:54.035 --> 00:02:58.410 The first is that constraints are easier for the model to satisfy when they are still in working memory. 72 00:02:58.410 --> 00:03:03.810 A banned-word list at the top of the prompt is something the model can check itself against, sentence by sentence, as it drafts. 73 00:03:03.810 --> 00:03:08.135 The same list at the bottom is something the model encounters after it has already written a draft in its head. 74 00:03:08.135 --> 00:03:13.847 It can self-correct, but self-correction is slower and less reliable than writing inside the lines from the start. 75 00:03:14.135 --> 00:03:17.160 You can feel the difference if you watch streaming output. 76 00:03:17.160 --> 00:03:19.310 Constraint-first prompts start clean. 77 00:03:19.310 --> 00:03:21.960 Constraint-last prompts hesitate. 78 00:03:22.235 --> 00:03:26.785 The second is that role and context, despite what most guides will tell you, are not constraints. 79 00:03:26.785 --> 00:03:27.460 They are setting. 80 00:03:27.460 --> 00:03:29.635 They tell the model the world the answer lives in. 81 00:03:29.635 --> 00:03:34.810 If you put them first, you are telling the model to optimize for fidelity to a persona, and persona is a soft target. 82 00:03:34.810 --> 00:03:41.772 The model will hit it, but it will hit it loosely, and any specific rules you bolt on afterward get filtered through that persona's defaults. 83 00:03:42.060 --> 00:03:48.835 The persona for "senior content strategist" includes the words "unlock" and "seamless," whether you like it or not. 84 00:03:48.835 --> 00:03:52.110 You have to outrank the persona, not append to it. 85 00:03:52.385 --> 00:03:55.510 The way to outrank a persona is to put the rules above it. 86 00:03:55.510 --> 00:03:57.310 That is the whole trick. 87 00:03:57.560 --> 00:04:03.010 The structure I use now I have a single template I run for every client deliverable. 88 00:04:03.010 --> 00:04:05.560 It has three sections and a fixed order. 89 00:04:05.560 --> 00:04:07.334 I do not deviate from the order. 90 00:04:07.334 --> 00:04:10.310 I deviate from the content all the time. 91 00:04:10.585 --> 00:04:16.135 The thing I want you to notice is that the "context and voice" section, which most prompt guides put at the top, is at the bottom in my version. 92 00:04:16.135 --> 00:04:17.947 That is not because context does not matter. 93 00:04:17.947 --> 00:04:18.697 It matters a lot. 94 00:04:18.697 --> 00:04:22.235 It is because context is a thing the model can use loosely without breaking the deliverable. 95 00:04:22.235 --> 00:04:23.097 Constraints are not. 96 00:04:23.097 --> 00:04:26.310 If the headline is 84 characters, the client cannot use the headline. 97 00:04:26.310 --> 00:04:30.685 If the voice is slightly off, the client can usually live with it or send a one-line note. 98 00:04:31.010 --> 00:04:34.610 You are ordering by what is non-negotiable, top to bottom. 99 00:04:34.610 --> 00:04:36.785 Put the things you will not edit at the top. 100 00:04:36.785 --> 00:04:39.510 Put the things you might edit at the bottom. 101 00:04:39.785 --> 00:04:41.010 A note on what this is not. 102 00:04:41.010 --> 00:04:44.910 This is not "the model cannot follow instructions at the end of a prompt." It can. 103 00:04:44.910 --> 00:04:51.135 Opus 4.7 is genuinely good at honoring long instruction sets in any order, and on a generous brief you will not see the difference. 104 00:04:51.135 --> 00:04:55.935 You will see the difference on tight briefs, on first drafts, on anything where you do not have time for a second pass. 105 00:04:55.935 --> 00:04:59.010 Which, if you are a working operator, is most things. 106 00:04:59.285 --> 00:05:03.010 What to do on Monday Pick one prompt you run regularly. 107 00:05:03.010 --> 00:05:07.897 Could be a meeting summary prompt, a draft-the-LinkedIn-post prompt, an audit prompt, whatever. 108 00:05:07.897 --> 00:05:12.460 Find the one where you most often have to do a second pass to fix the same kind of mistake. 109 00:05:12.460 --> 00:05:14.585 That repeated mistake is your constraint. 110 00:05:14.585 --> 00:05:19.285 It is the thing the model keeps forgetting because you keep putting it at the bottom of the prompt. 111 00:05:19.585 --> 00:05:20.460 Move it to the top. 112 00:05:20.460 --> 00:05:22.472 Make it the first thing the model reads. 113 00:05:22.472 --> 00:05:27.335 Run the prompt again with the same input you used last week and compare the two outputs side by side. 114 00:05:27.335 --> 00:05:32.485 If you do not see a difference, your prompt is already in good shape and you should be writing this article instead of reading it. 115 00:05:32.485 --> 00:05:38.110 If you do see a difference, congratulations, you just got a meaningful upgrade for the price of a copy-paste. 116 00:05:38.435 --> 00:05:41.385 The order is the cheapest optimization in prompting. 117 00:05:41.385 --> 00:05:42.410 It costs nothing. 118 00:05:42.410 --> 00:05:43.510 It ships today. 119 00:05:43.510 --> 00:05:46.760 You do not need a new model, a new tool, or a new framework. 120 00:05:46.760 --> 00:05:49.210 You need to move three lines up. 121 00:05:49.770 --> 00:05:50.620 That's the story. 122 00:05:50.620 --> 00:05:53.532 You'll find every source linked at the end of the article. 123 00:05:53.532 --> 00:05:56.195 Thanks for listening, on Promptway.