When I started this site, I wrote a lot about prompts. Specific prompts, prompt patterns, prompt engineering. Then I started doing real work with real people, and the prompts started mattering less.
What started mattering more was the context. What is in the model's window. What the model "knows" right now, in this turn, about the question. The difference between a good and a bad answer to the same question is almost always in the context, not in the question.
The phenomenon
Here is what I see, over and over. Someone has a long thread with a model. They have been refining their question, adding constraints, pushing back on bad answers. They ask me to look at it. I scroll to the first message and the problem is obvious: the first message was missing a fact that they only mentioned in message 17. By the time the model "understood", the answer was already in the noise of 16 other messages.
If they had started a new chat with the same fact moved to message one, they would have had the right answer in three turns instead of seventeen.
Why this is hard to teach
Because it is invisible. A bad prompt is visible: you can see the words, you can rewrite them. A bad context is invisible until the answer comes back wrong, and even then it is hard to attribute the failure to "the fact that this was message 14 instead of message 1".
The mental model that helps: treat every chat as a one-shot. Write the first message as if it were the only message. If the model can answer well from that message alone, you have done your job. If it can't, the next 20 messages are patches over a missing setup.
A small test you can run today
Take any of your long work chats. Identify the 3 most important facts in the final answer. Were they in the first message? If not, restart the chat and put them in the first message. The new answer is usually better, and arrives in fewer turns.
This is the skill. Prompting is the wrapper. The thing inside the wrapper is context.