Package Structure 01_paper Contains the final paper. 02_prompt_specification Contains the structured prompt specification and the exact prompts used to generate five Galaxy and five Nextflow workflows. 03_generated_workflows_and_validation Contains all generated workflow artefacts and validation materials in a single directory. The Galaxy artefacts are provided as .ga files. The Nextflow artefacts are provided as compressed DSL2 workflow projects containing the workflow code, configuration files, parameters, software environments, example inputs, helper files, and execution documentation where applicable. This directory also contains the workflow-specific validation reports and the manual expert-assessment results. Workflow Generation The workflows were generated using OpenAI's GPT-5.6 Thinking model through the ChatGPT interface between July 1 and July 23, 2026. Each task was conducted in a separate interaction to avoid conversational carryover. The model was not provided with the corresponding expert-curated workflow, and the generated artefacts were retained without manual correction. Assessment Two authors independently assessed the generated workflows and resolved disagreements through discussion. The assessment considered requirements coverage, analytical completeness, tool or module suitability, dataflow correctness, parameter completeness, platform conformity, and uncertainty handling. Requirements coverage, analytical completeness, tool or module suitability, parameter completeness, and uncertainty handling were assessed as satisfied. Dataflow correctness and platform conformity were partially satisfied because Galaxy W4 and Nextflow W3 require platform-specific revisions. Limitation The workflows were not executed in live Galaxy or Nextflow environments. The package supports reproduction of the prompt-based generation and manual-assessment procedures but does not establish confirmed end-to-end execution or scientific correctness.

This replication package accompanies the paper Prompt Engineering for Scientific Workflow Development. Large language models (LLMs) offer new opportunities to support scientific workflow development by translating natural-language descriptions into structured computational artefacts. However, workflow generation requires an LLM to interpret scientific objectives, identify analytical stages, select suitable tools, configure parameters, establish valid data dependencies, and follow the conventions of the target workflow system. Underspecified prompts may therefore produce incomplete analyses, inappropriate assumptions, invalid connections, or platform-incompatible representations. This paper presents a workflow-oriented structured prompt specification that explicitly represents scientific intent, data context, analytical requirements, tools and parameters, dataflow constraints, expected outputs, platform requirements, validation criteria, and unresolved information. We first identify six categories of information that workflow prompts should communicate and organise relevant prompting techniques into four groups: specification, decomposition and reasoning, grounding, and output control and refinement. We then apply the specification to ten expert-curated bioinformatics tasks spanning Galaxy and Nextflow. Independent assessment by two authors indicates that the generated artefacts satisfy requirements coverage, analytical completeness, tool or module suitability, parameter completeness, and uncertainty handling. Dataflow correctness and platform conformity are partially satisfied because two workflows require platform-specific revisions. These findings provide preliminary evidence that the specification can support the generation of coherent, implementation-oriented workflow artefacts while making missing assumptions visible. However, live execution testing and external workflow and domain-expert assessment remain necessary before deployment.