The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted natural language from users to generate campaign journey intents, creating a potential injection surface. This is effectively mitigated through a mandatory evidence chain:
Ingestion points: User natural language input processed in SKILL.md.
Boundary markers: The skill enforces a strict multi-step interaction flow requiring four pieces of confirmed information before proceeding.
Capability inventory: The skill calls internal tools (save_flow, query_cluster_qp_skill, query_channel_list) to create persistent canvas data.
Sanitization: All user input is translated into a fixed, validated JSON schema rather than being executed directly.
[COMMAND_EXECUTION]: The skill orchestrates the creation of marketing journeys by calling backend tools. These operations are restricted to the authorized projectId context provided by the platform and require explicit project selection from the user.
[REMOTE_CODE_EXECUTION]: No remote scripts, external package installations, or arbitrary shell commands were identified. Functional dependencies are managed through internal platform APIs.
[DATA_EXFILTRATION]: Network activity is restricted to internal tool calls. The skill generates platform-relative links (e.g., /#/hermes/flow/detail) for navigation, ensuring user data remains within the intended project environment.

journey-intent-parser