Instruction as compression
When you write a prompt, you compress a rich, private intention into a short public string. The model never receives the intention. It receives the string, and it has to — using nothing but the statistics of how words like yours have been used before.
Most of what we call a “wrong answer” is really a decompression error. In one review of failed requests, the majority traced not to the model’s reasoning but to an instruction that under-specified the goal — the words were followed faithfully; they simply weren’t the words the writer meant.
That reframes the skill. Prompting well is less about clever phrasing and more about :
The audience the output is for, not just the topic.
What a good answer would let you do next.
The failure you are most worried about.
A model does not know what you meant. It knows what you said, and what people who said things like that usually wanted. The gap between those two is the whole job.
Close that gap and the same model gets dramatically better — no new weights required. The intent gap is upstream of the intelligence, which is oddly reassuring: it is the part you can actually control.