1d • AI / Prompting
Making AI Create Not AI Slop Captions
I've recently been working on creating a system for content creation for my gym because I need structure and a way to it every time that's as frictionless as possible, otherwise I won't do it.
Captions are notoriously bad an noticeable for AI generation. They all sound the same.
"Not X. Not Y. Just Z."
Groupings in 3s
Em Dashes EVERYWHERE
and so on.
I had remembered seeing from Leila Hormozi maybe? a post or email about how you can spot AI language and there was some specific terms in there "formulaic structure", transition words, rhythmic and perfect structure.
So, I looked into it definitionally, what are the defining characteristics of AI generated captions that we can so EASILY tell are painfully AI?
I came up with some and tried playing around with some prompting. Here are the highlights and a sample prompt you can use and adjust based on your needs:
-
Write a caption for the following post. Before you output anything, apply every rule below without exception.
-No em dashes. If you write one, stop and rewrite that sentence as two sentences or restructure it entirely before continuing.
-Do not mirror the grammatical shape of consecutive sentences. "Not because X, but because Y" is a violation. Two sentences in a row that echo each other structurally is a violation. Stacked fragments, "Not X. Not Y. Just Z.", are a violation. If two consecutive sentences rhyme in structure, rewrite one of them.
-Do not group ideas in threes. If you notice a set of three parallel points, cut one or break the pattern.
-Cut all transition words. However, additionally, furthermore, therefore, ultimately, essentially. The ideas connect without signposting or they get restructured until they do.
-No dramatic one-line paragraphs. A single sentence standing alone for effect is an AI writing pattern. Fold it into surrounding prose.
-No formulaic pivots. "But here's the thing." "Here's the truth." "But here's what you need to know." Cut them. Make the point directly.
-The hook states something true. It does not ask a question. Stating something specific, slightly uncomfortable, or surprising stops people. A question as a hook is the most overused pattern in social media writing.
Before you output: read the draft once. Find every sentence that does not hold meaning on its own. Cut it. Then ask if the point could be made in fewer sentences without losing anything. If yes, rewrite. Then output.
-
Full context: I used Claude and was getting pretty good outputs. Still not perfect, depended on the question, information it had, resources it pulled from. But I also had some dramatically better ones. Even the bad ones we're much better than anything I was getting before.
I think a lot of the success and consistency comes from language during the instructions of "if you write one, stop and rewrite that sentence..." because getting fully rid of em dashes has been next to impossible to me, this has done a pretty good job so far.
And then I've been using this a lot on my prompting lately: "Before you output" and giving specific, reread, find X and replace with Y, rewrite, etc. It's been successful for me lately.
This was super helpful for me, you can probably use a lot of the same principles and rules on other prompting and generation stuff.
0
0 comments
Making AI Create Not AI Slop Captions
powered by
skool.com/exercise-programming-8878
Better programming, systems, and individualization for coaches and semi-private gym owners using modern tools without turning coaching generic.
Suggested communities
Powered by