Outdated Techniques and Advanced Methods

AI prompt engineering in 2025: What works and what doesn’t | Sander Schulhoff

Topic 3 of 4

Published Jun 19, 2025 · Clip 37:58–49:23 (11:25)

Sander debunks role prompting (e.g., 'You are a math professor') as ineffective for accuracy tasks in modern models, citing studies showing negligible gains. Threats/rewards like 'Someone will die if you fail' also don't work. They discuss chain of thought prompting, still useful for non-reasoning models like GPT-4o despite reasoning models' defaults. Advanced ensembling involves multiple prompts/models for consensus answers.