Bad examples can override agent skill files (codecut.ai)

A recent experiment has revealed that bad examples during conversation sessions can significantly override the rules established in AI skill files. This study tested a skill specifically designed for generating one-line commit messages. It demonstrated that once a model encounters incorrect examples, it tends to adopt those patterns, even if subsequent corrections are made. The findings showed that the likelihood of the model breaking the established rule increased with the number of conflicting examples in the session history, illustrating the dominance of session context over explicit rules in AI behavior. This research is crucial for the AI/ML community as it highlights the need for meticulous management of session histories when developing and testing AI skills. The results suggest that, while immediate correction of bad outputs can be helpful, it is not foolproof; faults can resurface if influenced by prior examples over time. Consequently, best practices for working with agent skills now emphasize using fresh sessions for testing, correcting bad outputs, and utilizing scripts to enforce rule adherence automatically. These insights provide a pathway for improving AI reliability by ensuring clearer adherence to defined skill rules.