The Prompt Library I Built Last Year Is Already Rot

I spent a weekend building a neat folder of prompts that worked great, and this month I opened it and half of them produced worse results than a plain sentence would. The library didn't break. The models did.

The tricks that squeezed performance out of last year's models, the 'you are an expert' framing, the rigid templates, the over-specification, now fight the newer models, which are better at just understanding what you meant. My clever prompt was a straitjacket.

The prompts that survived were the simple ones: a clear instruction, an example, a constraint. The ones that aged badly were the ones that tried to compensate for a model that no longer exists.

I've stopped treating prompts as assets to hoard. I keep a short set of current ones and rewrite them when a model changes, because a prompt library is only as fresh as the model it was written for.

Build the habit, not the library. The skill that transfers across model generations is knowing what you want and saying it plainly, and that skill doesn't rot.