Dan McKinley: Your Prompts Are Vectors and the Text Doesn't Matter
Dan McKinley gave a talk called Prompts Aren't Real and the thesis is exactly as blunt as the title. Prompt engineering as a craft is a category error. The prompts are vectors whose textual contents do not matter at all, and the only thing that does matter is the measurement apparatus you build around them. It is up at https://evaluation.club and took 91 points on Hacker News.
The evidence he leads with is the kind every team has and nobody writes down. A recurring JSON hallucination failure got fixed by renaming a field from title to heading. That is it. Nothing about the instruction changed, no reasoning was clarified, a word was swapped and the bug went away. If your mental model is that you are explaining a task to a colleague, that fix makes no sense. If your mental model is that you are searching a space and the text is a coordinate, it makes perfect sense and it also tells you that a human hand-picking coordinates is the wrong tool.
What he proposes instead is pass^k. Generate adversarial scenarios, run each of them many times, and establish a distribution before and after any change rather than a single passing run. Then let an algorithm with an LLM judge rewrite the prompt against that measurement, and accept that you will land somewhere very different from where you started and will not be able to explain why it is better in English.
Then the recursion bites. Judging something like brand voice is itself a fuzzy task, so you have to build and optimize an evaluator for it, which gives you an optimization problem nested inside your optimization problem. He does not pretend this is solved. He also notes that even at good model quality the demons escape containment once you watch at any real scale, and simple constraints like a character limit fail some fraction of the time no matter what the prompt says.
The uncomfortable implication for anyone hiring right now: the skill with market value is not writing the prompt, it is owning the loop that throws your prompt away.
← Back to all articles
The evidence he leads with is the kind every team has and nobody writes down. A recurring JSON hallucination failure got fixed by renaming a field from title to heading. That is it. Nothing about the instruction changed, no reasoning was clarified, a word was swapped and the bug went away. If your mental model is that you are explaining a task to a colleague, that fix makes no sense. If your mental model is that you are searching a space and the text is a coordinate, it makes perfect sense and it also tells you that a human hand-picking coordinates is the wrong tool.
What he proposes instead is pass^k. Generate adversarial scenarios, run each of them many times, and establish a distribution before and after any change rather than a single passing run. Then let an algorithm with an LLM judge rewrite the prompt against that measurement, and accept that you will land somewhere very different from where you started and will not be able to explain why it is better in English.
Then the recursion bites. Judging something like brand voice is itself a fuzzy task, so you have to build and optimize an evaluator for it, which gives you an optimization problem nested inside your optimization problem. He does not pretend this is solved. He also notes that even at good model quality the demons escape containment once you watch at any real scale, and simple constraints like a character limit fail some fraction of the time no matter what the prompt says.
The uncomfortable implication for anyone hiring right now: the skill with market value is not writing the prompt, it is owning the loop that throws your prompt away.
Comments