The Agent Skills Directory

[SAFE]: The skill performs legitimate research automation tasks using local scripts. The provided Python tools (discover_results.py and generate_results_table.py) use standard libraries for file operations and data processing with no evidence of network activity, credential harvesting, or persistence mechanisms.
[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection through the processing of untrusted data files.
Ingestion points: The agent reads data from paper/results/*.csv and notes/design/experiment-matrix.csv to update the paper's narrative and generate an abstract.
Boundary markers: There are no explicit boundary markers or instructions to the agent to disregard potential commands embedded within the CSV data.
Capability inventory: The agent has the ability to modify LaTeX source files (main.tex), execute local Python scripts, and run the paper compilation pipeline.
Sanitization: While the table generation script escapes special LaTeX characters for the source code, there is no validation or sanitization to prevent natural language instructions within the CSVs from influencing the agent's reasoning during the writing phases.

results-backfill