Mark Torres

I study how AI and social media algorithms interact with human psychology and sociology to shape social learning online. As a research engineer, I build large-scale data systems, AI-driven simulations, and LLM-based methods to understand and redesign digital environments for healthier discourse. I also consult with teams on shipping real-world AI systems as well as building AI agents and tools that drive value.

1 minute read

Quick tips that help me work better with LLMsPermalink

1. Use a Prompt RewriterPermalink

  • Claude and OpenAI have great prompt rewriters, such as Claude’s Prompt Improver.
  • These tools help refine your prompts for better outputs

2. Verify with Online ExamplesPermalink

  • When you doubt what ChatGPT says, ask it to find examples online and prove itself
  • More often than not, it corrects its own analysis if online resources don’t back it up
  • Even if it doesn’t self-correct, you get ground truth examples

3. Challenge Your BiasesPermalink

  • Ask ChatGPT (or any chat LLM) how your bias or beliefs could be wrong
  • LLMs are biased towards affirming what you already believe
  • Asking them to point out how you could be wrong is tremendously helpful

Side note 1: Take output from one LLM and pass it to another LLM asking how it could be wrong (Grok is great for this). Then pass both outputs to a third LLM and ask it to be a judge and give a final synopsis.

Side note 2: Ask LLMs to give 3 reasons pro/against a particular take.

4. Explore Conditional ScenariosPermalink

  • Besides giving a recommendation, ask ChatGPT under what circumstances or changes would its recommendation change
  • This way you get both a recommendation and know what would have to be different for its advice to change

5. Start Fresh ConversationsPermalink

  • Start a new chat window if you want ChatGPT to view your problem with fresh eyes
  • If you need context from a previous conversation, ask ChatGPT to summarize and compress your conversation so you can copy and paste it as context to the next conversation

6. Define Terms Very SpecificallyPermalink

  • Define terms very very specifically
  • Don’t assume the LLM knows what you mean by vague terms
  • Be explicit about definitions, especially for technical or domain-specific concepts

7. Add ExamplesPermalink

  • Add examples instead of using vague descriptors
  • Instead of saying “good writing”, add specific examples like “in the style of a journalist reporting on tech earnings” or “in the style of a NYTimes opinion piece”
  • Concrete examples help the LLM understand exactly what you want

8. Embrace Context EngineeringPermalink

  • The more you embrace context engineering, the better chance the LLM can do its job right
  • Use personas to guide the LLM to view a problem from a very specific point of view
  • Include only relevant context
  • Start new conversations if the current one has gone for too long

You may also enjoy