Prompt Engineering Approaches and System Prompts

Prompt Engineering

Mick McQuaid
mcq@utexas.edu

University of Texas at Austin

03 Mar 2025

Week Eight

Agenda

  • Presentations: Kylie, Zeel, Maitreyee
  • News
  • Review whatialreadyknow (Ishwari)
  • m2 questions
  • Prompt engineering techniques
  • System prompts
  • Sagemaker
  • Colab
  • Work time

Presentations

News

The Batch

⟨\langle pause to look at this week’s edition ⟩\rangle

WhatIAlreadyKnow (Ishwari)

Prompt Engineering Techniques and Goals

Prompt Engineering Techniques

  • Meta Prompting
  • AutoPrompt
  • Automatic Prompt Engineer
  • Gradientfree Instructional Prompt Search
  • Prompt Optimization with Textual Gradients
  • RLPrompt
  • Dialogue-comprised Policy-gradient-based Discreet Prompt Optimization

Tree of Prompt Engineering Techniques from Sahoo et al. (2024)

Tree of Prompt Engineering Techniques with top level branches listed on the following two frames

Goals of Prompt Engineering, Sahoo et al. (2024), 1 of 2

  • New tasks without extensive training
  • Reasoning and logic
  • Reduce hallucination
  • User interaction
  • Fine-tuning and optimization
  • Knowledge-based reasoning and generation
  • Improving consistency and coherence

Goals of Prompt Engineering, Sahoo et al. (2024), 2 of 2

  • Managing emotions and tone
  • Code generation and execution
  • Optimization and Efficiency
  • Understanding User Intent
  • Metacognition and Self-Reflection

Automatic Prompt Optimization, Wan et al. (2024)

  • Prompt engineering can be automated through APO
  • Two approaches:
    • Instruction Optimization
    • Exemplar Optimization
  • They are synergistic, but
    • EO is easier and more effective
    • Should reuse model-generated input-output pairs

System Prompts, Zhang et al. (2024)

What system prompts are

  • System prompts are general instructions that are not task specific but are included in a prompt
  • Some authors call them meta-instructions
  • An example might be “Let’s think step by step”
  • Recent studies show that the state-of-the-art prompt optimizer is ProTeGi
  • Accordingly, Zhang et al. (2024) test a system prompt optimizer with it and find that the system prompt optimizer performs better, noting that system prompts are more generally applicable

Zhang et al. (2024), continued

  • Zhang et al. (2024) cite recent research showing that combining multiple benchmarks in a single evaluation improves efficiency and alignment with human preferences, so they evaluate on a lot of task types
  • Zhang et al. (2024) uses a genetic algorithm to iterate over system prompts (tens of thousands at each step), relying on removing mistakes from output at each iteration

Example of successful system prompt

Write an answer that makes the reader feel happy. Write like you are explaining. First establish the set of facts you know, then answer the question based only on those facts.

Ethical considerations

(We covered this previously, but a reminder doesn’t hurt)

Admirably, Zhang et al. (2024) includes a section on ethical considerations. These include:

  • energy consumption of the process
  • possibility of stereotyping people, e.g., “act like a professor”
  • crowdsourced annotations from WEIRD contexts
  • need for research on model performance in new cultural and social contexts

Sagemaker Studio Lab

⟨\langle pause to look at the mistral tutorial ⟩\rangle

Google Colab

⟨\langle pause to demo Colab ⟩\rangle

END

References

Sahoo, Pranab, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. 2024. “A Systematic Survey of Prompt Engineering in Large Language Models: Techniques and Applications.” https://arxiv.org/abs/2402.07927.
Wan, Xingchen, Ruoxi Sun, Hootan Nakhost, and Sercan O. Arik. 2024. “Teach Better or Show Smarter? On Instructions and Exemplars in Automatic Prompt Optimization.” https://arxiv.org/abs/2406.15708.
Zhang, Lechen, Tolga Ergen, Lajanugen Logeswaran, Moontae Lee, and David Jurgens. 2024. “SPRIG: Improving Large Language Model Performance by System Prompt Optimization.” https://arxiv.org/abs/2410.14826.

Colophon

This slideshow was produced using quarto

Fonts are Roboto Light, Roboto Bold, and JetBrains Mono Nerd Font