Why it matters: Piling more requirements onto a Claude prompt does not produce more control, it produces a model that silently resolves your contradictions and burns its reasoning budget on parsing instead of thinking.
The Details:
When a prompt carries many constraints that pull in different directions, the model makes an invisible tradeoff. Ask for something extensible and under fifty lines at once, and it picks one or splits the difference into a version that satisfies neither, with no signal to you that a choice was made at all.
Every extra clause draws from the same cognitive account as the actual problem-solving. A model spending its effort decoding a fifteen-point spec has less left to notice a cleaner path, or to flag that two of your own requirements conflict three lines apart.
Short prompts fail too, but differently. A one-line request for a cache produces something functionally correct and completely irrelevant to the real bottleneck. The fix in both directions is the same missing ingredient: intent. State what you are trying to achieve and the one or two facts about the system the model could not infer from the code, and stop there.
The pattern reverses for code review. Generation rewards terse intent; review rewards a bit more context. A model told to check for concurrency issues catches something real. A model given only "review this" defaults to style nitpicks because it has nothing to weigh risk against.
Bottom Line: The best prompt is rarely the longest one you can write or the shortest one you can get away with. It is the shortest one that still says why.
Enjoy this article?
Listen to the Claude Code Conversations radio show or join the community.
