The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches and processes large volumes of external text from user-provided or search-discovered URLs.
Ingestion points: Web content is scraped by the scripts/fetch-sources.py script and then read in full by the agent in Step 4 using cat -n.
Boundary markers: The skill provides a structured markdown template and explicit instructions to quote sources verbatim, which helps differentiate between user input and external data, though it does not fully prevent instruction override from malicious content.
Capability inventory: The agent can perform web searches via deepapi, write analysis files to the local file system, and execute the provided Python scraping script.
Sanitization: The fetch-sources.py script includes a clean function that uses regular expressions to strip layout junk (tables, nav bars) and normalize whitespace from the scraped content.
[COMMAND_EXECUTION]: The skill relies on a local Python script (scripts/fetch-sources.py) that uses the subprocess module to execute system commands.
The script calls curl to interact with the scraping API, bypassing potential limitations of Python's standard HTTP libraries in certain environments.
[DATA_EXFILTRATION]: The skill's scraping mechanism involves sending URLs and authentication headers to an external API (API_BASE_URL). While this is functional, it involves the transmission of potentially sensitive metadata (URLs of interest) to a remote service.
[EXTERNAL_DOWNLOADS]: The workflow involves downloading content from arbitrary external websites and a specific scraping API. The integrity and safety of the content fetched from these URLs are not verified beyond basic text cleaning.

signal-from-expert