What You Prompt is What You Get: Increasing Transparency of Prompting Using Prompt Cards

Based on the paper "What You Prompt Is What You Get: Increasing Transparency Using Prompt Cards" by Caut et al., here is a detailed technical summary covering the problem, methodology, contributions, results, and significance.

1. Problem Statement

The rapid proliferation of Large Language Models (LLMs) has established prompt engineering as a critical discipline for guiding model outputs. However, the field faces significant challenges regarding reproducibility, transparency, and evaluation:

• Lack of Standardization: Unlike traditional machine learning (ML) where models are trained and documented, prompts are often ad-hoc, complex, and lack standardized documentation.
• Reproducibility Issues: LLM outputs are highly sensitive to prompt phrasing, ordering, context, and inherent stochasticity (randomness). Reproducing results is difficult due to missing details on datasets, model versions, and contextual constraints.
• Evaluation Limitations: Standard ML benchmarks (e.g., accuracy, AUC) are often ill-suited for prompt engineering. Many real-world tasks (like transforming data into narrative text) do not fit into existing benchmark frameworks. Furthermore, benchmark data contamination (models training on test sets) undermines validity.
• Ethical and Safety Risks: Prompts can inadvertently amplify biases, hallucinate facts, or leak sensitive data, yet these risks are rarely documented systematically.

2. Methodology

The authors propose a new documentation framework called Prompt Cards, adapting the concept of Model Cards (introduced by Mitchell et al., 2019) specifically for the domain of prompt engineering.

The Prompt Card Framework:
The paper defines a structured set of sections to document the entire prompt engineering lifecycle. Key sections include:

• Model Details: Specific LLM version, license, and training parameters (crucial as LLMs update frequently).
• Intended Use: Explicit scope of the prompt (e.g., educational vs. professional) and limitations.
• Dataset: Documentation of input data sources, construction methods, and inherent biases.
• Context: Detailed description of system prompts, few-shot examples, external knowledge injection (RAG), and normative models used to shape the output.
• Prompt Architecture: A schematic breakdown of the prompt structure (e.g., "Tell it who," "Tell it what," "Tell it how").
• Factors: Demographic and environmental variables affecting performance (e.g., specific population groups covered by the data).
• Quantitative Analysis: Instead of broad benchmarking, this section focuses on hypothesis-driven evaluation (e.g., testing if a prompt faithfully reconstructs data) using methods like "LLM-as-a-judge" or human evaluation.
• Data Security: Assessment of privacy risks, internet connectivity requirements, and strategies for anonymizing sensitive inputs.
• Ethical Considerations: Explicit discussion of bias, fairness, cultural universality, and potential for misuse.
• Caveats & Recommendations: Known failure modes and guidance for responsible deployment.

Case Study Application:
To validate the framework, the authors applied Prompt Cards to a task called "Wordalisation" (transforming structured numerical data into natural language text). They demonstrated this across three applications:

1. Football Scout: Converting player statistics into scouting reports.
2. World Value Survey: Transforming survey data into country profiles.
3. Personality Test: Converting Big Five personality scores into candidate summaries.

3. Key Contributions

• Conceptual Innovation: The introduction of the Prompt Card as a standardized artifact to bridge the gap between prompt engineering and scientific reproducibility.
• Framework Adaptation: Successfully adapting the Model Card framework to address the unique challenges of in-context learning and prompt-based systems, distinguishing it from traditional model training documentation.
• Evaluation Paradigm Shift: Arguing against the reliance on universal benchmarks for all prompt tasks. Instead, the paper advocates for contextual, hypothesis-driven quantitative analysis combined with qualitative transparency.
• Practical Implementation: Providing concrete examples of how to document sensitive issues like data privacy (e.g., anonymizing raw scores before prompting) and ethical biases (e.g., cultural limitations of personality tests).
• Open Resources: The authors have made the prompt cards and the associated applications available on GitHub and a web interface for public reference.

4. Results and Findings

• Feasibility of Documentation: The authors successfully documented complex prompt engineering pipelines (including statistical preprocessing, normative modeling, and few-shot prompting) using the Prompt Card structure.
• Quantitative Validation: In the World Value Survey example, they used an "LLM-as-judge" approach to test the faithfulness of the generated text. By reconstructing original data categories from the generated text, they measured reconstruction accuracy, demonstrating that quantitative metrics can be used for specific hypothesis testing rather than general benchmarking.
• Identification of Hidden Biases: The documentation process forced the identification of implicit assumptions, such as the cultural bias in the Big Five personality test and the demographic limitations of the football scout dataset (male-only, specific league).
• Security Awareness: The framework successfully highlighted security risks, such as the potential for data leakage when using third-party LLMs, and proposed mitigation strategies like local preprocessing.

5. Significance and Impact

• Enhancing Reproducibility: Prompt Cards provide a "step-by-step guide" that allows researchers and practitioners to replicate prompt-based experiments, addressing a major gap in current AI research.
• Regulatory Alignment: The framework aligns with emerging AI regulations (e.g., EU AI Act, US AI Initiative) that demand transparency, risk assessment, and auditability. It serves as a bridge between regulatory intent and technical implementation.
• Responsible AI Deployment: By making design decisions, ethical limitations, and failure modes explicit, Prompt Cards empower users to make informed decisions about when not to use a system, reducing the risk of harm in high-stakes domains (healthcare, legal, hiring).
• Paradigm Shift in Evaluation: The paper challenges the "benchmark-obsessed" culture of NLP, suggesting that for many complex, real-world tasks, transparency and documentation are more valuable than a single numerical score. It draws a parallel to data journalism, where the process of storytelling and data handling is as important as the final output.

In conclusion, the paper argues that as prompt engineering becomes a core component of AI systems, the industry must move beyond treating prompts as "black box" inputs. The Prompt Card offers a structured, transparent, and ethical mechanism to document, evaluate, and govern these systems, ensuring they are used responsibly and reproducibly.