Stop prompting, start designing
At a recent meetup, a developer told me why he had given up on AI assistants. They always agreed with him, so one day he instructed his to disagree. From then on it disagreed with everything. Both versions were useless, and for him that settled the question: the tools are not good enough.
I think that story is the whole problem in miniature, because the instruction was the mistake, not the model. "Disagree with me" is a posture command. A model can only obey it the way it obeyed the previous default posture, and so you trade a sycophant for a contrarian parrot. What he actually wanted was judgment. Judgment does not come from commanding a personality. It comes from context: what you are trying to achieve, what your constraints are, what good looks like, what has already been decided. Give a model that, and disagreement shows up when it is warranted, which is the only kind worth having.
Here is the analogy that makes it click. Imagine you hire someone, and on their first day you give them nothing. No documentation, no introduction to the project, no access to the tools, no explanation of how the team works. Then you fire questions at them and argue with the answers. When they underperform, nobody blames the hire. Everybody blames the onboarding. We have known this about people forever; poor onboarding reliably tanks the performance of even excellent hires.
Now notice that most people use AI exactly like that. A blank chat window, a question, a fight about the answer, repeat. No standing instructions, no context about the project, no definition of what a good result looks like, no memory of past corrections. And then the conclusion is that the hire is not good enough.
Working well with AI is not a prompting trick. It is a design discipline. You are not asking questions, you are building the system the conversation happens inside: durable instructions that encode how you work, defined processes and skills for recurring tasks, feedback that turns into standing rules instead of evaporating, guardrails around what it can touch and what it must never touch. The mental model is onboarding a very fast new hire, not querying an oracle.
I have seen the click happen in real time. Talking with skeptics, arguments got me nowhere until I showed the actual architecture of my setup: the instruction files, the defined skills, the guardrails, the context the agents work against. The reaction was immediate: oh, now I see what you mean. And the honest part of the demo is that it took days to design. The value was never in the asking. It was in the system that makes the asking work.
There is a reason this gap exists, and it is worth saying plainly: this is a management skill, and most engineers have never managed anyone. Delegating well, briefing well, building an environment where someone else can perform, giving feedback that becomes durable. People who have led teams recognize the whole playbook instantly. People who have only ever executed keep looking for a better prompt.
So if AI keeps disappointing you, the question to ask is not whether the model is good enough. It is whether you have built anything for it to work inside. Stop prompting. Start designing.