The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data from Craft documents, creating a potential surface for indirect prompt injection.
Ingestion points: scripts/fetch.py fetches content from public craft.me share links.
Boundary markers: SKILL.md includes explicit instructions for the agent to treat document content as data and not execute embedded instructions.
Capability inventory: The skill uses subprocess.run to call curl for network requests but does not perform file writes.
Sanitization: No explicit sanitization of document content is performed before rendering to Markdown, relying on the agent's instructions.
[COMMAND_EXECUTION]: The script executes shell commands via the subprocess module.
Evidence: scripts/fetch.py uses subprocess.run(["curl", ...]) to fetch data.
Mitigation: The command uses a list of arguments (avoiding shell interpolation) and the target URL is strictly validated to ensure it belongs to the craft.me domain.
[EXTERNAL_DOWNLOADS]: The skill fetches data from an external service.
Evidence: scripts/fetch.py downloads share JSON from craft.me APIs.
Context: The resource is fetched from a well-known document service, and access is restricted to public share identifiers via regex validation.

read-craft