What's in a Soul? A Provenance-Complete Dataset of AI Agent Persona Prompts in the Wild on GitHub

AI agents are increasingly defined by documents rather than code. Files such as SOUL.md, persona.md, and character cards declare an agent's identity, values, tone, and behavioral limits, and they are accumulating in public repositories. Prior work that mined prompts from repositories studied strings embedded in source code or the management practices around prompt files, but not what these files declare about the agent. We release a dataset of 281 persona prompts from 168 public GitHub repositories, collected with five path-based searches and a manual inclusion rule. Every record pins repository, path, and commit SHA (281/281) and was re-verified byte-for-byte at that SHA. Because 42% of source repositories carry no license, a redistribution gate implemented as record fields ships full text for the 179 records (64%) whose licenses permit it and metadata plus SHA for the remaining 102. An accompanying two-axis labeling scheme codes content (nine binary facets and a count of conditional clauses) separately from form (an ordinal prose/flat/full-pack scale). In independent full coding by two raters, form was reliable (Krippendorff's α = 0.84) while content facets fell below threshold (α from −0.11 to 0.55). Disagreements were mostly one-directional, so the raters placed the boundaries of concepts such as identity and goal in different places. We release both raters' labels and the disagreement lists. Under both raters, behavioral rules (83–90%) and conditional instructions (72–84%) dominate, and form skews to full-pack (74–78%), with files that declare nothing concentrated outside it. The soul path, the most common by repository count, has the most uniform writing style. The dataset supports persona-fidelity stimuli, classification-boundary studies, and longitudinal tracking of conventions. DOI: 10.5281/zenodo.23230144.