The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and summarize large volumes of untrusted data from external web sources. This creates a surface for indirect prompt injection where instructions hidden in the sources could influence the AI's behavior.\n
Ingestion points: references/sources.md identifies multiple external data sources including web pages (via Firecrawl or Jina Reader), Reddit, Hacker News, YouTube, and arXiv.\n
Boundary markers: The prompt template in templates/digest_prompt.md uses a simple <items JSON> placeholder to insert data into the prompt without explicit instructions to the model to ignore any embedded commands within that data.\n
Capability inventory: As described in SKILL.md, the skill writes code to the local file system, executes that code to process data, and registers persistence via scheduled tasks.\n
Sanitization: The documentation does not specify any sanitization or filtering of the fetched content to remove potential injection strings before passing them to the LLM.\n- [COMMAND_EXECUTION]: The core functionality involves generating Python scripts and executing them in the local environment to perform data fetching, storage, and processing tasks.\n- [PERSISTENCE]: The skill sets up automated execution using system tools like cron, Windows Task Scheduler, or macOS launchd. While intended for the daily digest feature, this represents a persistence mechanism on the host system.\n- [EXTERNAL_DOWNLOADS]: The skill encourages the use of third-party APIs and scrapers (Apify, Firecrawl, Jina Reader, OpenRouter) and specific community scripts (e.g., harshmaur/reddit-scraper) to acquire data.

build-signal-engine