You'd think a more detailed prompt would be a more effective prompt.
Ask a 70-billion-parameter model where Alice is standing relative to Charlie, given that Alice is north of Bob and Bob is west of Charlie.
Gartner declared prompt engineering dead in July 2025. Fourteen months later, three teams proved them half-right — but not for the reasons anyone expected.
You've probably done this: a model gives a wrong answer, so you increase the thinking budget. More tokens, more reasoning, surely it'll get there.
Most of the prompts in production right now were written for models that needed help thinking.
The conventional wisdom has been dead for months, and most prompt engineers haven't noticed. An EMNLP25 paper from Cheng et al.
Sergey Brin told people to threaten their models. So Wharton's AI lab ran the experiment.
You read the reasoning trace. It looks perfect — the model identified the key constraint, considered two approaches, picked the right one, and explained why.
I pulled up a client's system prompt last week. Fourteen lines, every one starting with "Do not.
Most teams I talk to treat their JSON schema as plumbing. The prompt gets iterated on for weeks.
Ask a model to write a function, then ask it to check the function for bugs. Nine times out of ten: "The implementation looks correct.
A team I talked to last month was spending $2,000 per developer per month on Claude API calls.
Most people think of chain-of-thought as the model explaining its work.
Last month I spent two days tuning a classification prompt. Rewrote the system message maybe fifteen times.
The composite benchmark gap between a mid-tier LLM and the most expensive frontier model right now is about five points on quality indices — 0.75 versus 0.
I spent three days last month rewriting a system prompt for a code-review agent. Tried persona frames, XML structure, numbered constraints, the works.
Every prompt engineering tutorial from the last three years drilled the same lesson: show the model what you want.
I audited a scoring pipeline last week that was sampling every request twenty times and taking the majority vote.
Something about prompt engineering has been bugging me. We tell people "make the model think step by step" as if that's always the right advice.