For years now, advice on prompt engineering has often sounded pretty comical: “Tell the model it is a brilliant expert”, “Promise a $100 tip for cooperation”, “Tell the model it will go to jail if it disobeys you”.
These tips were certainly effective, not because they created human-like motivations in the model (as far as we know?), but rather because they keyed into low-level associations between words and phrases that the model learned during training.
Researchers at UPenn have now studied these and other techniques using the best present-day models (Meincke et al. 2025). One of their core findings: the colorful tricks just don’t matter anymore. Advancements in late-stage fine-tuning have led to models that are so good at following instructions that you needn’t try to manipulate them with promises or threats.
I would venture that the precise format of prompts matters less and less as well, for the same reason that the colorful tricks lost their effect: model improvements.
So: certain kinds of prompt engineering are no longer needed. Excellent news! However, it is vital not to over-interpret this as saying that prompts don’t matter. Prompts matter more than ever before. This is precisely why the model providers have invested so heavily in instruction-following: so that you can be successful writing complex prompts that contain intricate sets of instructions.
Your focus now should be on making sure that your prompts actually contain all the necessary information. This is a collaborative, creative process of discovering what counts as good. In general, even subject matter experts are not able to comprehensively list all their requirements right off the tops of their heads. They need to react to diverse examples, interact with colleagues, check external documents, see test results, etc.
You do still need to get all this information into the prompt in a way that coherently defines the behavior you want to see from your system. Advancements in prompt optimization can automate this for you. The human process of discovery is much more significant.
At Bigspin, our sole focus is ensuring that you succeed in this complex process of shaping GenAI systems. Whether you are starting from scratch or improving an existing system, we make it easy for you to collect feedback on examples, interact with colleagues, run evaluations, and partner with the Bigspin AI itself to expose weaknesses and find edge cases. In the end, we will turn this into an optimal prompt for you. (I confess that I am a bit sad that these optimized prompts will no longer begin with colorful phrases like “You are a brilliant expert”.)
