When does adding more instructions to a prompt start to HURT rather than help?

Adding instructions feels like it should always help. Every rule covers one more case. In practice there is a clear point where more instructions start to hurt. The reason is attention. The model has to spread its attention across every directive in the prompt. With three or four sharp rules, each one gets weighted strongly and the model follows them. Past roughly seven to ten core directives, attention thins out and the model starts selectively following some while quietly ignoring others. Behavior gets inconsistent. Token cost also climbs linearly. You pay for every instruction on every call, whether the model is following it or not. So past the directive budget, you are paying more for worse compliance. The practical move when a prompt has grown to fifteen or twenty rules: consolidate redundant directives into one, prioritize the ones that matter most, then trim everything else. Fewer, sharper rules beat a long list.

Prompts have a directive budget. Past roughly seven to ten core instructions, compliance plateaus or regresses while token cost keeps climbing linearly. The mechanism is attention competition: every directive has to fight for weight in the model's attention pattern, and with too many directives, some get selectively ignored. The failure mode is the long checklist prompt. A team adds one rule per bug report, the system prompt grows to twenty bullet points, and the model starts ignoring half of them. The team's instinct is to add more rules to fix the new failures, which makes the problem worse. The fix is a three-step pass. Consolidate redundant directives into one rule with a worked example. Prioritize the directives that map to the most failures. Trim everything else, including rules that were added once and never moved a metric. A tight ten-directive prompt usually beats a sprawling twenty five directive one on the same eval set. Context window is rarely the binding constraint. The directive budget is hit long before token limits.