The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes a dedicated Python script scripts/material-prep.py to handle the extraction and chunking of large source materials. This script uses subprocess.run to invoke pdftotext for PDF processing. The implementation adheres to security best practices by passing arguments as a list and avoiding shell execution, which prevents shell injection vulnerabilities. Execution is further restricted by platform-level allowed-tools configuration.
[EXTERNAL_DOWNLOADS]: To support distillation from web-based sources, the material-prep.py script fetches content from remote URLs using urllib.request. It includes specific logic to handle GitHub blob URLs by converting them to raw content links, ensuring that only document data is retrieved for processing.
[DATA_EXFILTRATION]: The skill includes proactive measures to prevent the unintended ingestion of sensitive data. The helper script maintains a comprehensive blocklist of sensitive directory names (e.g., .ssh, .aws, .git) and file patterns (e.g., *token*, *secret*, .env). These locations are skipped during recursive directory walks unless the user explicitly overrides the behavior with the --include-sensitive flag.
[PROMPT_INJECTION]: Because the skill processes untrusted external content like PDFs and web pages, it is exposed to indirect prompt injection risks. Malicious instructions hidden in source documents could attempt to hijack the distillation process. The skill mitigates this through explicit instructions to the agent to paraphrase and extract actionable rules and workflows rather than performing verbatim copying, reducing the likelihood of executing embedded commands.

aif-distillation