Things to squash before your message 'stands out'｜Three outputs that will cause accidents if sent as-is, discovered by actually running reply prompts

In the previous article, I considered why leaving reply drafting entirely to ChatGPT makes it 'stand out' (sound unnatural). The cause is not the AI's performance, but the lack of information provided. The relationship with the recipient, the temperature of the conversation so far, and the rules implicitly present in the situation—if you have the AI write without providing these three things, the text will sound out of place.

Even if you understand the reason, if you don't know what to instruct and how, you'll end up having to rewrite everything from scratch. 'I understand I need to convey the relationship. But specifically, with what wording?' That is likely the hurdle you face after reading the previous post.

Therefore, this time, I created instruction prompts for internal chat, external email, and customer support, and actually sent them to a generative AI.

The result revealed something different from what I had anticipated.

Before worrying about whether it sounds 'out of place' or not, there were multiple outputs that would cause practical accidents if sent as-is. This is what needs to be squashed before adjusting the writing style.

Specifically, there are three. Returning a polite message without pointing out contradictions in the instructions. Creating numbers or conditions that were not provided. Promising things that were not agreed upon.

All of these are natural as sentences, and because they are neatly polished, they are hard to notice before sending. From here on, I will present the actual instruction prompts, the outputs returned, the points where I stumbled, and how I corrected them in order.