Ruben Hassid · 5 min read · Gen AI Tools
Prompt Engineering 101.
Read the original on ruben.substack.com →
Prompt engineering for LLMs evolves quickly, making older tactics obsolete. The piece advocates consulting fresh research and distills a few robust rules for crafting prompts—focusing on neutral questions, avoiding step-by-step gimmicks, and guarding against over-trusting model confidence. It warns about hallucinations and the risks of heavy rule-loading, urging a simple, goal-driven approach.
Key Takeaways
- Never end your prompt with right? Use a neutral question instead.
- Forget the step-by-step approach; a concise prompt plus a role can be more effective.
- Do not trust the model's confidence; verify references and open sources.
- Stick to at most three rules in a prompt; simplicity improves accuracy.
- Prioritize explicit goals over long lists of examples; align prompts to the task.
Excerpt
There is no training on how to ask AI correctly. And AI changes so fast that techniques from before don’t work anymore. LLMs (like ChatGPT or Claude) changed so much it’s outdated. For example, I used to tell everyone to add “ Take a deep breath and work on this step by step ” to their prompt. Because science said so: “ Large Language Models as Optimizers ” (also known as the OPRO paper) But this paper is from September 2023 and talks about the model PaLM 2-L from Google. You now have access to Claude & ChatGPT models that…