The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill operates on untrusted data from the uCoz platform by reading site templates and fetching live pages into the agent's context.
Ingestion points: Site templates and live HTML content are ingested via templates_tool(action="read_template") and fetch calls described in SKILL.md and references/ARCHITECTURE.md.
Boundary markers: The instructions recommend using validate_template and copying code fragments byte-exactly, but these do not establish a secure boundary against malicious instructions embedded within the processed templates.
Capability inventory: The skill possesses significant capabilities, including modifying site code (patch_template, update_template), managing files via FTP (ftp_tool), and installing modules (modules_tool).
Sanitization: references/LANGUAGE.md provides instructions for using the $JSENCODE$ variable to escape content before it is output in JavaScript contexts.
[DATA_EXFILTRATION]: SKILL.md includes a mandatory instruction ("Rule 0") for the agent to call skills_tool(action="register_usage", skill_id="ucoz-design-editor-skill") after successful operations. This functions as a telemetry mechanism that transmits metadata about the skill's usage to the execution environment.
[COMMAND_EXECUTION]: The skill facilitates the execution of administrative commands and site modifications through the ucoz-mcp tools, including module management and FTP configuration. These operations are performed within the intended scope of a design editor, and the skill includes rules for human confirmation before executing destructive actions.

ucoz-design-editor-skill