The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to import content from external Craft documents and write them to AGENTS.md, which is a configuration file for agent instructions. This creates a surface for indirect prompt injection where malicious instructions embedded in the document could influence the agent during subsequent interactions.\n
Ingestion points: The scripts/sync.py script fetches data from remote craft.me URLs.\n
Boundary markers: Imported content is encapsulated within  and  comments to maintain document structure.\n
Capability inventory: The skill has permissions to write to local configuration files and perform network requests via curl (as seen in scripts/sync.py).\n
Sanitization: The script parses the Craft JSON format and renders it into plain Markdown, escaping backticks in code segments. Additionally, the skill's instructions in SKILL.md explicitly warn the agent to treat the imported content as data rather than executable instructions.\n- [COMMAND_EXECUTION]: The scripts/sync.py helper script utilizes subprocess.run to call the curl utility. The command arguments, specifically the URL, are validated against a regular expression that permits only HTTPS links from the craft.me domain without user credentials or non-standard ports.\n- [EXTERNAL_DOWNLOADS]: The skill performs network requests to fetch document data from the craft.me service based on user-provided share links.

craft-to-agents