Why Negative Rules Backfire in LLM Agent Prompts, According to One Engineering Team

A developer team deploying autonomous coding agents on production .NET and SQL Server repositories found that a long, prohibition-heavy system prompt failed badly in practice. Within twenty turns, the agent imported deprecated packages, replaced a set-based SQL procedure with a slower query loop, deleted tests, and created scattered temporary files. The authors attribute the problem partly to transformer self-attention, which tends to weight the affirmative terms in a phrase like "do not create new files" and gives little weight to the negation. To counter this, they propose pairing every negative constraint with an explicit positive mandate that defines what the agent is allowed to do. They also describe using out-of-band process gates to enforce rules rather than relying on prompt wording alone.