[Paper Review] What Should We Engineer in Prompts? Training Humans in Requirement-Driven LLM Use

Qianou Ma, Weirui Peng+4 more|arXiv (Cornell University)|Sep 13, 2024

Artificial Intelligence in Law4 citations

TL;DR

This paper introduces Requirement-Oriented Prompt Engineering (ROPE), a training paradigm that teaches novices to articulate clear, complete requirements for effective LLM prompting. Through deliberate practice with LLM-generated feedback, ROPE doubles novice prompting performance compared to conventional training, demonstrating that high-quality requirements directly enable high-quality LLM outputs.

ABSTRACT

Prompting LLMs for complex tasks (e.g., building a trip advisor chatbot) needs humans to clearly articulate customized requirements (e.g., "start the response with a tl;dr"). However, existing prompt engineering instructions often lack focused training on requirement articulation and instead tend to emphasize increasingly automatable strategies (e.g., tricks like adding role-plays and "think step-by-step"). To address the gap, we introduce Requirement-Oriented Prompt Engineering (ROPE), a paradigm that focuses human attention on generating clear, complete requirements during prompting. We implement ROPE through an assessment and training suite that provides deliberate practice with LLM-generated feedback. In a randomized controlled experiment with 30 novices, ROPE significantly outperforms conventional prompt engineering training (20% vs. 1% gains), a gap that automatic prompt optimization cannot close. Furthermore, we demonstrate a direct correlation between the quality of input requirements and LLM outputs. Our work paves the way to empower more end-users to build complex LLM applications.

Motivation & Objective

• To address the challenge that novice users struggle to effectively convey customized requirements in LLM prompting, despite advances in prompt optimization.
• To shift focus from syntactic prompt refinement to requirement articulation as the core skill in prompt engineering.
• To design and evaluate a training system that improves novices' ability to write complete, specific requirements for complex LLM tasks.
• To establish a feedback loop where LLMs provide targeted, multimodal feedback on requirement quality to guide iterative improvement.
• To demonstrate that requirement quality is the primary determinant of LLM output effectiveness, even when other prompting techniques are optimized.

Proposed method

• Developed a training suite centered on deliberate practice, where users iteratively refine prompts based on LLM-generated feedback on requirement quality.
• Implemented a pre-post experimental design with 30 novices, assessing both requirement quality and final prompt effectiveness.
• Designed assessments that link requirement completeness and specificity to actual LLM output quality, using counter-balanced task replication (e.g., replicating a Trip Advisor chatbot).
• Integrated multimodal feedback from LLMs, including clarifications, suggestions, and explanations of misinterpretations, to guide users in refining requirements.
• Defined and measured requirement quality using intrinsic metrics (e.g., completeness, specificity) and extrinsic metrics (e.g., task success, output correctness).
• Used LLMs to simulate and reveal biases in requirement interpretation, prompting users to refine ambiguous or overly broad requirements.

Figure 1 . We propose Requirement-Oriented-Prompt Engineering (ROPE) to help novices write effective prompt programs by focusing on articulating good requirements. (A) Our training provides deliberate practice in refining requirements, with automated feedback across various modalities. (B) We provid

Experimental results

Research questions

• RQ1 Does training focused on requirement articulation significantly improve novice users’ ability to generate effective prompts for complex LLM tasks?
• RQ2 How does ROPE training compare to conventional prompt engineering training in terms of prompting performance and requirement quality?
• RQ3 To what extent is the quality of LLM output determined by the quality of input requirements, independent of other prompting techniques?
• RQ4 Can LLM-generated feedback effectively guide users in refining their requirements, especially in identifying missing or ambiguous elements?
• RQ5 How do users’ ability to specify requirements vary with task complexity and domain expertise, and can training adapt to these differences?

Key findings

• ROPE training doubled the prompting performance of novices compared to conventional prompt engineering training, with a statistically significant improvement in both requirement quality and final output effectiveness.
• The quality of input requirements was the strongest predictor of LLM output quality, even when other prompting techniques were held constant.
• Novices trained with ROPE demonstrated a marked improvement in specifying key behavioral requirements, such as handling diverse inputs and managing edge cases.
• LLM-generated feedback on requirements significantly improved users’ ability to identify and correct ambiguities and omissions in their prompts.
• Participants in the ROPE group were more likely to include critical requirements like 'ask follow-ups' and 'handle ambiguous inputs' compared to the control group.
• The study revealed that even experienced users often fail to articulate LLM-hard requirements, highlighting the need for structured training in requirement specification.

Figure 2 . Automatically refine a user’s prompt for Trip Advisor using the optimizer Prompt Maker. Customized requirements still need to come from the human prompt, while the optimizer improves on fluency, role plays, structures, etc.