K-Dense-AI/scientific-agent-skills
Scientific Agent Skills test: analyze GitHub Trending with a reproducible workflow
After skipping an already completed #1 project, we selected the #2 Trending repository and tested two of its 163 skills on a real 20-repository snapshot.
01 · PROBLEM
What I wanted to solve
A successful bulk installation did not show whether any scientific workflow could complete a real task, preserve a failure, produce reusable artifacts, and state limitations without overstating the other untested skills.
02 · PROCESS
What I did
We installed only statistical-analysis and scientific-visualization, preregistered a two-sided Spearman test, kept all 20 Trending repositories, and ran 10,000 bootstrap iterations. The first report export failed because an optional dependency was missing, so we replaced that export step without adding a package and reran the analysis.
03 · RESULT
What actually happened
The second run exited with code 0. Spearman rho was 0.0451 with p=0.8502 and a 95% bootstrap interval from -0.3983 to 0.4754. The experiment delivered the raw and enhanced CSV files, script, JSON summary, PNG, SVG, export manifest, run log, report, article, and video package.
04 · PITFALLS
What others should watch for
This is a 20-repository one-day sample selected by the Trending mechanism. It cannot describe all of GitHub, establish causality, measure repository quality, or predict future growth. The remaining 161 skills were not individually executed.
05 · VERDICT
My conclusion
The statistical-analysis and scientific-visualization workflows completed a real task and left an auditable evidence chain. Use the repository selectively, install only the skills needed for the task, and review each skill's code, network, package, database, model, and domain requirements.