The Agent Skills Directory

[PROMPT_INJECTION]: The skill architecture is susceptible to indirect prompt injection as it processes untrusted data from arbitrary URLs and interpolates it into LLM prompts.
Ingestion points: External data is ingested in scraper/sources/source_name.py via requests or playwright.
Boundary markers: The prompt in ai/pipeline.py uses markdown headers (e.g., # Items, # User Context) to separate data from instructions, though it lacks explicit 'ignore embedded instructions' warnings.
Capability inventory: The agent's capabilities include writing to a Notion database and updating a local feedback.json file. It does not possess RCE or system-level write capabilities based on LLM output.
Sanitization: The scraped content is not sanitized before being formatted into the prompt, which is a common characteristic of data-enrichment agents.
[EXTERNAL_DOWNLOADS]: The skill utilizes standard, well-known Python packages including requests, beautifulsoup4, playwright, and notion-client. These are fetched from official package registries.
[DATA_EXFILTRATION]: The skill is designed to send scraped content and user-defined context to Google's Gemini API for enrichment and to the Notion API for storage. Both are well-known, trusted services, and the data transfer is the intended core functionality of the agent.

data-scraper-agent