No dataset. I'm using two types of prompts to direct the LLM on outputs, an overall system prompt that applies to all the outputs and eight style prompts corresponding to each of the eight cringe categories.
You're right, repetition can be an issue. I have anti-repetition logic on the opening sentences (it passes recent openers back in and tells the model to avoid them), but nothing on reused metaphors and analogies. And there are definite grooves the model likes to default to (the number 47, for instance, and people named Gerald).
Hello, I'm the creator of LinkedIn CringeBot 3000.
The "Oops..." error message some of you were getting was because my Anthropic credit ran out in the middle of the night my time and I needed to refill it. We should be good going forward.
CringeBot uses Claude Sonnet 4.6 as the main LLM (though after you have generated an output you can hit the "Try DeepSeek" button to see how DeepSeek would answer the same prompt).
One interesting wrinkle. As far as I can tell, Sonnet 4.6 is the last Anthropic model to offer a "temperature" setting. Temperature controls how random or predictable the model's word choices are, which is useful for tuning the outputs. Newer Anthropic models don't seem to have this option.
There are two prompt categories at work under the hood. A general system prompt that applies to all the outputs and 8 separate "style prompts" that correspond to the 8 different cringe styles. Refining these has been the hardest part of the project.
