The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The viewer.html file fetches the SheetJS library from cdn.sheetjs.com to enable spreadsheet rendering within the evaluation viewer. This is a well-known technology service used for document processing.
[COMMAND_EXECUTION]: Several scripts, including run_eval.py and improve_description.py, utilize the Python subprocess module to invoke the platform's pi CLI for running evaluations and generating descriptions. The generate_review.py script also uses system commands like lsof and kill to manage local network ports for its HTTP server.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it processes user-provided test cases and feedback to refine instructions.
Ingestion points: Untrusted data enters the context via evals.json and feedback.json, which are read during the improvement cycle.
Boundary markers: The skill uses XML-style delimiters (e.g., <current_description>, <skill_content>) to separate instructions from processed data.
Capability inventory: Subprocess execution capabilities are present in run_eval.py and improve_description.py to facilitate testing.
Sanitization: The scripts use regular expressions to strictly extract content within expected tags, reducing the risk of the model following instructions embedded within the data.
[DATA_EXPOSURE]: The skill manages data within a local workspace directory, reading transcripts and saving feedback. This data handling is confined to the local filesystem and the local web server provided for the user.

skill-creator