DELTA 

This repository contains supplementary materials for the research paper: "Can Large Language Models Decompose User Stories into Tasks? Exploring the Role of Prompting Strategies and Models". The paper was submitted to the Automated Software Engineering Conference 2026.

Structure

• DELTA - Pipeline implementation for decomposing user stories into tasks and evaluating the quality of the resulting decompositions
• Prompt iterations – Prompt engineering experiments and observations
• Descriptive statistics – Task title length statistics
• Evaluation session – Human evaluation session including informed consent form and evaluation session protocol
• Statistics – Per research question (RQ) statistical results

Research Questions

RQ1
How do persona-based zero- and few-shot prompting strategies affect text similarity between LLM-generated and human-created decompositions?

RQ2
How do model family and size affect text similarity between LLM-generated and human-created decompositions?

RQ3
To what extent do text similarity metrics align with human expert judgments of decomposition quality?

RQ4
To what extent does the criteria-based automated evaluation module of DELTA agree with human expert judgments of decomposition quality?