The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external CLI tools including defuddle and firecrawl via shell commands to parse web content.
Evidence: defuddle parse "<url>" --markdown and firecrawl scrape "<url>" --only-main-content in references/extract-article-news.md.
Mitigation: The skill includes explicit instructions in SKILL.md for the agent to reject URLs containing shell metacharacters (quotes, backticks, $, ;, |, &, <, >, newlines) and to pass URLs as single quoted arguments to prevent injection.
[INDIRECT_PROMPT_INJECTION]: The skill fetches content from user-provided URLs and processes it using natural language instructions defined in templates.
Ingestion points: Untrusted content is fetched from external URLs (articles, social threads, etc.) via tools like Firecrawl, TinyFish, and Defuddle.
Boundary markers: The skill checkpoints raw content into extraction.json and stages a draft.md before publication, though it does not specify explicit LLM-level delimiters for the untrusted content within the prompt context.
Capability inventory: The skill has the ability to write files to the /using-slipbox directory and execute local CLI tools.
Sanitization: The skill implements a /write-checks gate for synthesized content (quoted instructions) which applies 'Style' and 'Humanize' checks, and it enforces strict URL parsing to prevent shell-level exploits.
[EXTERNAL_DOWNLOADS]: The skill interacts with external services to fetch data.
Evidence: Uses Firecrawl for scraping and the youtube-transcript-api to fetch video data from YouTube. These are well-known services for web data extraction.

clip-resource