The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands using placeholders {{DEFUDDLE_PYTHON}} and {{DEFUDDLE_SCRIPT}}. These commands incorporate user-controlled inputs via the --url and --project flags. While the instructions advise the agent to treat --project as an untrusted path and check for path traversal characters like .., this pattern still constitutes a potential command injection surface.
[EXTERNAL_DOWNLOADS]: The skill depends on an external integration referenced as spaceage64/claude-defuddle. It also notes a reliance on third-party utilities such as yt-dlp and pandoc for processing media and documents.
[PROMPT_INJECTION]: The skill is designed to ingest data from external sources, such as web pages and media transcripts, which are outside the user's direct control. This exposes the agent to Indirect Prompt Injection (IPI).
Ingestion points: The --url parameter in SKILL.md triggers the retrieval of external content.
Boundary markers: The agent is instructed to verify the content (checking title, canonical URL, and completeness) before using it to answer questions.
Capability inventory: The agent can execute shell commands, perform network requests through the tool, and write files to a local Obsidian vault.
Sanitization: The skill includes specific instructions for the agent to manually validate project paths and verify extraction results.
[DATA_EXFILTRATION]: The skill documentation mentions ~/.claude/CLAUDE.md as a legacy configuration path for storing credentials. Accessing hidden configuration files within a user's home directory is a sensitive operation that could lead to unauthorized data exposure if not handled correctly.

claude-defuddle