The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted research paper text, PDFs, and reference lists across several prompts, creating a potential surface for indirect prompt injection. * Ingestion points: Paper text and reference strings are interpolated directly into prompts such as citation-f1-prompt.md and litreview-quality-prompt.md. * Boundary markers: While headers like 'Paper Text:' are present, the prompts lack explicit instructions for the LLM to ignore or escape instructions that may be embedded within the papers themselves. * Capability inventory: The skill executes a local Python script (scripts/compute_f1.py) and writes JSON reports to the file system. * Sanitization: There is no evidence of sanitization or escaping of the input content before it is processed by the model.
[COMMAND_EXECUTION]: The skill involves the execution of a local Python script, scripts/compute_f1.py, which calculates Precision, Recall, and F1 scores based on JSON data produced by the agent. The script uses standard libraries and is included as part of the skill package.

paper-autoraters