Tips on how to use GenAI efficiently

I want to share few tips on how to use GenAI for the researching and learning efficiently:

  1. Build agents to your tasks.
    • Agent is just an LLM with precise and detailed instructions containing: the role LLM takes, area of expertise, what to take into account, how to process information (what to emphasize, what to ignore), format of answers, structure, markup (if needed). Tip: Use standard LLM model helping you with making instructions for your agent.
    • Provide to your agent only tools it needs, nothing extra.
  • If it is possible - set up necessary level of agent “creativity” via temperature parameter: low temp → more accurate, less creative; higher temp → more creative, but also less accurate.

    • I have agents such as: “Generic helper”, “AWS DevOps helper”, “Linux administration helper”, etc.

  1. For repetitive tasks create parametrized prompts. E.g. I have a prompt which helps me to extract transcription from a given YouTube video and analyze it.
  2. Try to apply critical thinking instructions for a model in all relevant prompts. Directly ask model to evaluate statements and use critical thinking approach, if possible - verify used data, identify bias, incorrect argumentation and fallacies. But do not forget to instruct model to outline positive, well done, correct statements and arguments.
  3. Ask follow-up question to each specific point separately.
  4. Use branching of conversation.
  5. Be aware of the LLM context window. Different models have different context window sizes and the longer your conversation is, the higher chance is that model will start “forgetting” the beginning. It will impact quality of your research.
  6. Hence, if you need to dive deep into specific question - create new chat for it or use branching inside the same chat. I personally prefer creating new chats.
  7. Never fully trust model answer. LLMs do not count and LLMs do not know grammar, they see vectors of numbers.
    • Never trust model calculations if they are made with use of external tools (like calculator, generation and execution of a program, etc.)
      • If you need complex calculations, then ask models to use code.
      • Depending on the GenAI tool you use, the model can also call sandboxed environment to run generated code, and you can see result (or even an error) right away, in the chat.
      • Or you need to copy given code and run it yourself. Yes, you shall understand programming.
    • Never trust model grammar. If you want to be the most accurate, use grammar checking tools. The MS Word, Google Docs, LibreOffice Writer with grammar check will do.
    • Never trust fully that data, which model presented to you is correct. Always check the data.
  8. Be skeptical! Use your critical thinking and try to challenge models’ answers and arguments. If you see gaps - always ask for more data to prove/disprove a point. And don’t hesitate to use web-search yourself and verify data which model presented to you.

And one more thing…

GenAI is NOT an expert or a reliable source of information.

But GenAI IS a great training and sparring PARTNER, and it can HELP YOU to improve YOUR expertise, YOUR skill of argumentation, YOUR critical thinking.