The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from Slack messages, issue tracker comments, and shared documentation to synthesize rationale. This creates a surface for indirect prompt injection if malicious instructions are embedded in the analyzed data.
Ingestion points: Slack channels, Linear tickets, Notion pages, and Sentry issue logs via MCP tools.
Boundary markers: The instructions prioritize verbatim quoting and citations to anchor claims, but do not provide explicit sanitization or escaping mechanisms for the ingested content passed between subagents.
Capability inventory: The skill possesses read access to multiple enterprise data sources and executes local shell commands (git, gh).
Sanitization: No specific sanitization or filtering logic is mentioned for handling adversarial input within the source material.
[COMMAND_EXECUTION]: The skill uses shell commands including git blame, git log, and gh pr view to establish a context anchor. These commands use arguments (file paths, line ranges, commit hashes) derived from user queries and repository state.
[DYNAMIC_EXECUTION]: The skill implements a multi-agent orchestration workflow, dynamically spawning specialized 'investigator' subagents with custom prompts and model configurations at runtime based on the available MCP servers in the user environment.

why