By skyforbes My main points are:
Research shows that “bad prompt” can’t be defined. If one can’t define what’s bad, then no engineering is possible.
Tweaking phrasing wastes time compared to improving data quality, retrieval, and evaluations.
Prompt techniques are fragile and break when models get update. Prompts don’t work equally well across different models and even across different versions of the same model.
The space attracts grifts: selling prompt packs is mostly a scam and this scam inflated importance of the so-called engineering.
Prompts should be minimal, auditable, and treated as a thin UI layer. Semantically similar prompts should lead to similar outputs. The user shouldn’t be telling a model it’s an expert and not to hallucinate – that’s all just noise and a problem with transformers
Prompting can’t solve major problems of LLMs – hallucinations, non-determinism, prompt sensitivity and sycophancy – so don’t obsess with it too much.
Models don’t have common sense – they are incapable of consistently asking meaningful follow-up questions if not enough information is given.
They are unstable, a space or a comma might lead to a completely different output, even if the semantics stay the same.
The better the model, less prompting is needed because prompt sensitivity is a problem to solve and not a technique to learn.
All in all, cramping all possible context into the prompt and begging it not to hallucinate is not a discipline to learn but rather a technique to tolerate till models get better.
I would post the article with references to studies etc. but I feel like it might be not allowed. It is not hard to find it though.