The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions in references/data-collection.md to download files from arbitrary, user-defined or search-discovered URLs using the curl command.
[REMOTE_CODE_EXECUTION]: The instructions direct the agent to perform runtime installation of software packages and binaries using brew install (e.g., for poppler and pandoc) and pip install (e.g., for pdfplumber, python-docx, python-pptx) to handle file conversions.
[COMMAND_EXECUTION]: The skill requires the use of shell commands such as curl, pdftotext, and pandoc to process data from external sources. These commands are executed with parameters derived from external metadata (filenames, URLs), which may lead to command injection if not sanitized.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted information gathered from the web and external documents. 1. Ingestion points: Content from scraped web pages and text extracted from downloaded PDF, DOCX, and PPTX files in references/data-collection.md. 2. Boundary markers: The skill lacks instructions for using delimiters or warnings to ignore instructions embedded within the ingested data. 3. Capability inventory: The agent has access to web searching, scraping tools, local file system writing, and shell command execution across all analysis phases. 4. Sanitization: There are no documented procedures for sanitizing or filtering external content before it is incorporated into the final analysis reports.
[DYNAMIC_EXECUTION]: The workflow involves dynamically installing and calling external utilities to process binary document formats into text at runtime, creating a pathway for executing code based on external input.

competitive-analysis