The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from multiple external sources (GitHub PRs, Slack, Notion, Linear, etc.) and uses sub-agents to analyze this content. While it lacks strict boundary markers, it includes extensive instructions for 'epistemic discipline' to ensure objective reporting and prevent the agent from being misled by the data.
Ingestion Points: Slack threads, Notion pages, Linear tickets, and GitHub PR comments.
Capability Inventory: The skill uses sub-agents for reasoning and synthesis but is explicitly instructed not to perform write operations.
Sanitization: No explicit technical sanitization or input escaping is defined for the aggregated data.
[COMMAND_EXECUTION]: The skill executes shell commands such as git blame, git log, and gh pr view to retrieve metadata and history from the local repository and GitHub. Arguments for these commands are derived from file paths and commit hashes identified during the investigation.
[EXTERNAL_DOWNLOADS]: The skill interacts with well-known developer services (GitHub, Slack, Notion, Sentry, Datadog, Databricks, Linear) via Model Context Protocol (MCP) tools to fetch relevant context for its investigation. These interactions are aligned with the skill's primary purpose and target established, trusted platforms.

why