The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches content from external URLs provided by the user. If an attacker controls the target website, they could embed malicious instructions that the AI agent might follow when processing the extracted text.
Ingestion points: The defuddle parse <url> command in SKILL.md allows the agent to ingest content from arbitrary web pages.
Boundary markers: Absent. There are no instructions to the agent to treat the fetched content as data only or to ignore any instructions embedded within it.
Capability inventory: The skill invokes a CLI tool to read and parse web content.
Sanitization: The skill description states it removes "clutter" like ads and navigation, but it does not specify sanitization or filtering of prompt injection techniques within the remaining text content.
[EXTERNAL_DOWNLOADS]: The skill instructs the user to install an external package globally using npm install -g defuddle. This downloads and installs code from a public registry.
[COMMAND_EXECUTION]: The skill operates by executing the defuddle CLI tool through shell commands, passing user-provided URLs as arguments.

defuddle