Agents' skills are too long (mine were too) (aac.social)

Anthropic has made a significant move in optimizing their AI model, Claude Code, by cutting its system prompts by approximately 80%. In a recent interview, Thariq from the Claude Code team explained that as AI models evolve, they require less prescriptive guidance, which can negatively impact output quality. The excess detail in prompts stemmed from cumulative edits and feedback over time, leading to unnecessarily lengthy instructions. To address this, Thariq developed "skillmin," a skill minimizer that ensures any cuts made to the prompts maintain the model's performance and behavior through rigorous evaluation methods. The technical process behind skillmin involves two agents: one that creates an evaluation of the original skill and another that minimizes the skill while keeping essential commands intact. The eval compares the original and the minimized version side by side to ensure they behave identically. This approach underscores a broader implication for the AI/ML community: many existing AI skills may harbor unnecessary complexity that could be stripped away without sacrificing functionality. By streamlining these skills, the approach not only saves computation tokens but also enhances model efficiency. The minimized skills are model-agnostic and have already been successfully deployed, reinforcing the principle that less can often be more in AI instruction design.