The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill incorporates a 'Demand-signal' branch that ingests untrusted text from public forums, reviews, and social media threads, creating a surface area for indirect prompt injection attacks.
Ingestion points: External data is collected from platforms including Reddit, Hacker News, X (Twitter), and GitHub issues through discovery tools like last30days and social-fetch (references/demand-signals.md).
Boundary markers: The instructions mitigate risks by advising the agent to 'Quote minimally, paraphrase by default' and to strictly distinguish between observed evidence and internal inferences.
Capability inventory: The skill utilizes web-scraping and browser-automation capabilities (Firecrawl, Browserbase) to retrieve content from these external sources.
Sanitization: The skill includes explicit instructions to escape every value when generating standalone HTML evidence reports from collected data to prevent cross-site scripting or prompt confusion.

prospecting