The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a custom CLI tool (releases) to perform administrative actions such as fetching source updates (admin source fetch) and committing new overviews (admin overview update). It also coordinates batch processes by executing a local script via the Workflow tool.
[INDIRECT_PROMPT_INJECTION]: The skill involves a surface for indirect prompt injection because it processes untrusted release content from external sources (GitHub, web scrapes) to generate AI summaries.
Ingestion points: External data enters the context via the releases admin overview inputs command output.
Boundary markers: Absent. The prompt template for sub-agents does not use specific delimiters or instructions to treat the external release content as untrusted.
Capability inventory: Subprocess execution (releases CLI), file writing (/tmp/ and ~/.releases/work/), and agent spawning (Workflow).
Sanitization: The skill includes style linting and HTML entity de-escaping, but lacks security-focused sanitization to prevent adversarial instructions in the source data from affecting agent behavior.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill implements a systematic run-recording mechanism that writes audit logs, command history, and session traces to the user's ~/.releases/work/ directory for operational auditing and cost tracking.
[DYNAMIC_EXECUTION]: Complex batch maintenance is handled by executing a repository-local script (.claude/workflows/update-overviews.ts) to manage parallel sub-agent tasks and workflow logic.

maintaining-orgs