Hot take: my "prompt library" was just me copying my own mistakes
I spent about 4 months building a doc of 60 prompts I thought were gold, saved them all in Notion and reused them for everything. Then a buddy from my old job in Denver asked me why my outputs always sounded the same, kind of flat and stiff. I went back and read my prompts out loud and realized half of them just said "be helpful and accurate" with no real instructions, so the model was guessing the whole time. What tipped me off was testing 3 prompts side by side and getting nearly word for word the same answer. So how do you all write prompts that actually change the output instead of just dressing it up?
Watched a buddy go through almost the same thing last year. He had like 40 prompts saved for writing product descriptions and swore by them, until his boss told him every listing read like it came from the same robot. Turns out his prompts were all like "write a good description, make it sound nice," zero detail about tone, length, who was reading it, what made the product different. He ran an experiment where he fed three totally different products through the same prompt and got back three paragraphs that were basically twins. The fix for him was adding stuff like "write this for a 30 year old guy who hates ads" or "keep it under 50 words, no fluffy words." Once he did that, the outputs actually started sounding different from each other. The prompt isn't a magic spell, it's more like a job order, and if the order is vague the model just fills in the blanks with the same safe junk every time.