The Agent Skills Directory

[COMMAND_EXECUTION]: The skill directly interpolates the user-supplied variable $ARGUMENTS into shell commands, such as parallel-cli research run "$ARGUMENTS". An attacker could provide input with malicious quoting or additional flags to manipulate the command's behavior. Although the execution environment is restricted to the parallel-cli binary by the allowed-tools manifest, argument injection remains a risk.
[EXTERNAL_DOWNLOADS]: The documentation instructs users to update or install software via pipx upgrade parallel-web-tools and a platform-specific /parallel-setup command. These actions involve downloading and installing external code from the vendor's repositories.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted data retrieved from the internet.
Ingestion points: Web content is ingested via parallel-cli research run and parallel-cli research poll (Step 1 and Step 2).
Boundary markers: Absent. The skill does not implement delimiters or 'ignore' instructions for the data being researched.
Capability inventory: The skill has access to the filesystem (writing via -o in Step 2) and the parallel-cli toolset.
Sanitization: There is no evidence of sanitization for the executive summaries or markdown reports generated from external web sources before they are presented to the agent context.

parallel-deep-research