The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/bootstrap.sh script invokes multiple system package managers (apt-get, dnf, yum, pacman, brew, winget, choco) to install the Python 3 runtime environment. It specifically uses sudo -n to attempt non-interactive privilege escalation for these installations.
[REMOTE_CODE_EXECUTION]: scripts/clip_articles.py executes npx -y defuddle@0.19.2, which involves downloading and running code from an external registry (npm) at runtime.
[EXTERNAL_DOWNLOADS]: The skill communicates with several external APIs and services to extract article content, including api.fxtwitter.com, metaso.cn (Reader API), and r.jina.ai (Jina Reader).
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests untrusted content from the web and writes it to the local filesystem without robust content sanitization.
Ingestion points: Article content is fetched from arbitrary URLs via the fetch and parse_article methods in scripts/clip_articles.py.
Boundary markers: The output Markdown files use YAML frontmatter as a structural boundary, but provide no instruction-level isolation or explicit warnings for the agent regarding the article body content.
Capability inventory: Includes local file writing (Path.write_text), network requests (urllib.request), and system command execution (subprocess.run).
Sanitization: Content processing is limited to HTML entity unescaping and whitespace normalization, which is insufficient to prevent malicious instructions embedded in web pages from influencing the agent's behavior.

web-clipper