Thursday, January 25, 2024

Prompt patterns

 Just sharing a summary of  a paper that tried to develop a catalog of prompt patterns. The sourcez;

"A Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPT" by Jules White, Quchen Fu, Sam Hays, Michael Sandborn, Carlos Olea, Henry Gilbert, Ashraf Elnashar, Jesse Spencer-Smith, Douglas C. Schmidt. Arxiv. https://doi.org/10.48550/arXiv.2302.11382 

  1. Meta Language Creation Pattern: Focuses on creating a custom language for LLMs to improve their understanding of prompts.
  2. Output Automater Pattern: Aims to automate the generation of actionable steps or scripts in response to prompts.
  3. Flipped Interaction Pattern: Involves reversing the typical interaction flow, with the LLM posing questions to the user.
  4. Persona Pattern: Assigns a specific persona or role to an LLM to guide its output generation.
  5. Question Refinement Pattern: Enhances the LLM's responses by refining the user's questions for clarity and focus.
  6. Alternative Approaches Pattern: Encourages the LLM to offer different methods or perspectives for tackling a task.
  7. Cognitive Verifier Pattern: Involves the LLM generating sub-questions to better understand and respond to the main query.
  8. Fact Check List Pattern: Guides the LLM to produce a list of facts or statements in its output for verification.
  9. Template Pattern: Involves using a predefined template to shape the LLM's responses.
  10. Infinite Generation Pattern: Enables the LLM to continuously generate output without repeated user prompts.
  11. Visualization Generator Pattern: Focuses on generating text outputs that can be converted into visualizations by other tools.
  12. Game Play Pattern: Directs the LLM to structure its outputs in the form of a game.
  13. Reflection Pattern: Encourages the LLM to introspect and analyze its own outputs for potential errors or improvements.
  14. Refusal Breaker Pattern: Designed to rephrase user queries in situations where the LLM initially refuses to respond.
  15. Context Manager Pattern: Controls the contextual information within which the LLM operates to tailor its responses.
  16. Recipe Pattern: Helps users obtain a sequence of steps or actions to achieve a desired result.

Each pattern is detailed with its intent, context, structure, key ideas, example implementations, and potential consequences.

I want to acknowledge a good attempt, but am not sure this list is very intuitive or very helpful. In practical terms, we either ask questions or give tasks, defining some output parameters - like genre, audience, style, etc. However someone might find this helpful to keep thinking. We do need some way of classifying prompts. 

Prompt as a magic incantation

In engagements with AI, the crafting of prompts—a crucial interface between human intention and machine output—has acquired an almost mystic...