The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external sources including web search results, social media posts, and public professional profiles, creating a surface for potential indirect prompt injection attacks.
Ingestion points: Data is ingested via tools such as web_search_live, web_enrich_live, social_post_list_live, and person_search (LinkedIn data).
Boundary markers: The skill relies on structured data returns from execution scripts but does not explicitly use text-based delimiters to separate external content in the final prompt.
Capability inventory: The skill has access to JavaScript execution (execute tool), Python script generation, file system writes (HTML artifacts), and shell access for Chrome screenshots.
Sanitization: Instructions specify filtering and projecting data within the execution environment (e.g., using .map() and .slice()) to limit the content returned to the agent.
[DYNAMIC_EXECUTION]: The skill is designed around an execute tool provided by the Crustdata MCP server, which runs plain JavaScript for data transformation. It also instructs the agent to generate Python scripts to create HTML org chart artifacts.
Evidence: SKILL.md contains multiple JavaScript code snippets for data processing, and orgchart-playbook.md provides a methodology for dynamic Python script generation to handle hierarchy data.
[COMMAND_EXECUTION]: In development environments like Claude Code, the skill performs local shell operations to generate and review artifacts.
Evidence: orgchart-playbook.md details using the system's Chrome or Chromium binary to take headless screenshots of generated HTML files for self-review.
[EXTERNAL_DOWNLOADS]: The skill fetches images and data from external services to populate its research reports.
Evidence: Downloads company logos and person profile pictures which are then base64-inlined into standalone HTML files to ensure proper rendering.
[DATA_EXFILTRATION]: The skill has the capability to read sensitive data from connected internal tools such as CRM systems, email, and team chat.
Evidence: Instructions mandate that all internal data access is read-only and used solely to build account history for the user's research. No evidence was found of sensitive data being transmitted to unauthorized or third-party domains.

account-research