The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes large volumes of unstructured data from external sources such as Slack messages, PR comments, and Notion documents, which may contain instructions designed to influence the agent's conclusions.
Ingestion points: Data retrieved from MCP tools for Source Control, Issue Trackers, Documents, Chat, Observability, and Error Tracking platforms.
Boundary markers: The investigator and synthesizer prompt templates lack explicit delimiters or instructions to ignore embedded commands within the fetched data records, though they emphasize epistemic discipline.
Capability inventory: Subagents are granted access to MCP tools with 'readonly: false' (agent mode), enabling them to interact with various organizational APIs.
Sanitization: No explicit sanitization or filtering of external content is performed before processing by the LLM.
[COMMAND_EXECUTION]: The skill uses local and remote repository tools to gather initial context for investigations.
Evidence: Execution of git blame, git log, and gh pr view to establish the "Code Anchor" in Step 2 of SKILL.md.
[DATA_EXFILTRATION]: The skill is designed to aggregate and synthesize sensitive information from across a company's infrastructure and communication tools.
Evidence: Step 3 of SKILL.md directs agents to perform a "full parallel investigation" to ensure "complete coverage," accessing platforms like Datadog, Sentry, Databricks, Slack, and Notion to build a cited rationale.

why