PromptLab: The AI Agent That Tests What Actually Works on LLMs

Author tested whether threats, guilt, or flattery change LLM behavior when estimating unlabeled chart values; across 500 trials threats performed worse, flattery helped slightly, and results depended on model, task, and whether the prompt mentioned evaluation; conclusion: anecdotes overgeneralize and reproducible experiments are needed.