The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external sources (such as web scrapes, changelogs, and GitHub releases) via the releases admin overview inputs command. This content is then used by sub-agents to generate AI overviews, which introduces a risk of indirect prompt injection where malicious content in a release note could influence the agent. The skill includes mitigation steps such as a linting pass for prohibited editorializing and a mandatory HTML entity de-escaping process.
Ingestion points: Data retrieved through the releases admin overview inputs --json command in SKILL.md.
Boundary markers: The skill instructs sub-agents to return content in specific fenced markdown and JSON blocks and defines strict style rules.
Capability inventory: Access to the releases CLI, file system writes to /tmp/ and ~/.releases/, and the ability to spawn sub-agents.
Sanitization: Implements a linting pass to check for banned phrases and word limits, and performs HTML entity decoding to ensure citation offsets remain valid.
[COMMAND_EXECUTION]: The skill makes extensive use of the releases CLI, a vendor-specific tool from buildinternet, to perform administrative functions like fetching remote sources (admin source fetch), updating overviews (admin overview update), and managing local workspaces.
[DYNAMIC_EXECUTION]: The skill uses a Workflow harness to execute local TypeScript scripts (e.g., .claude/workflows/update-overviews.ts) and parallelizes operations by spawning multiple sub-agents using the agent() tool.

maintaining-orgs