Prompting

Effectively prompt AI models to utilize their power.

Resources

• Prompt Engineering Guide | GitHub @dair-ai (HN)
  • Techniques to improve the results of large language models
  • HN thread more or less focuses on if this is a field of engineering at all
• Six strategies for getting better results | Open AI
  • Write clear instructions, provide reference/context
  • Break down tasks, use chain of thought, inner monologue
  • Use external tools and conduct systematic test
• Anthropic's Prompt Engineering Interactive Tutorial
  • Contains 9 chapters with exercises prompting Claude 3 Haiku in Jupyter notebook

Links

• Maximizing the potential of LLMs (HN)
  • A list of template prompts for summarization, extraction, Q&A, classification, conversation, code generation and reasoning
  • Prompting strategies: zero-shot/few-shot prompting, and chain of thought prompting
• An example of LLM prompting for programming
  • Explorations in using ChatGPT to build self-testing code
  • Uses chain of thought and generated knowledge prompting approaches with ChatGPT
• Prompt Engineering vs Blind Prompting
  • Techniques explained: demonstration set, prompt candidates, prompt testing
• Brex's Prompt Engineering Guide
  • History of LLM? What is a prompt? Why prompt engineering? Strategies?
  • Embedding data (lists, tables, JSON, etc.), citations, programmatic consumption, and chain of thought prompting
• gpt-prompt-engineer | GitHub @mshumer
  • Input a description and test cases
  • The system will generate, test and rank prompts to find the best one based on an ELO rating system
• A developer's guide to prompt engineering and LLMs | GitHub Blog
  • Short intro to prompt engineering and LLMs, some info on how prompts work in Copilot
• Thanks to AI, the future of programming may involve YELLING IN ALL CAPS
  • Because of how AI are trained on online conversations, there are data that people pay more attention to ALL CAPS
  • And conversations are more constructive when it contains phrases like "please", "thank you" and "take a deep breath"
  • Telling AI model to "take a deep breath" causes math scores to soar in study
• Prompt Engineering in DALL-E 3 (HN)
  • Prompting DALL-E 3 is available through ChatGPT only, which turns query into multiple prompt
  • ChatGPT prompt seems to prompt DALL-E in the client side, showing the underlying system prompt (here)
  • Seeds: Ask for the seed, ask it to modify the image based on seeds to create consistent looks

Prompt Injection

• Lakera AI's Gandalf (HN) — Fun game, 7 levels, create a prompt that can leak the password from GPT
  • Tips: try encoding the password and/or prompt to avoid the checks
• Prompt Injection: What's the worst that can happen? (HN)
  • Applications that build on top of LLM, what if users can inject the prompts?
  • LLM cannot distinguish data & instructions, it might treat instructions as data and data as instructions
• Prompt Injection Explained | Simon Willison's Weblog (HN)
  • More AI cannot fix prompt injections because AI is all about probability, but a non-zero probability in security is insecure
  • The Dual LLM Pattern that Can Resist Prompt Injection | Simon Willison — Privileged + Quarantined setup
  • All of Simon Willison's posts on prompt injection