The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted natural language requirements and data models to generate API specifications, documentation, and project logs. This creates a vulnerability surface where malicious instructions could be embedded in input data to influence the behavior of the agent or downstream systems.
Ingestion points: The skill reads user requirements, data models (SCHEMA_TO_GATEWAY), and implementation constraints (BUILDER_TO_GATEWAY) across its SURVEY and DESIGN phases.
Boundary markers: While the skill uses structured templates, it does not explicitly instruct the agent to use delimiters or ignore embedded instructions within processed data to prevent obedience to malicious requirements.
Capability inventory: The skill has the capability to write design rationales and project updates to .agents/gateway.md and .agents/PROJECT.md on the host file system.
Sanitization: There is no evidence of specific sanitization or filtering logic applied to incoming data before it is interpolated into prompts or generated artifacts.
[OBFUSCATION]: The reference materials contain examples of Base64-encoded strings (e.g., 'eyJpZCI6MTIzfQ==') used as cursors for pagination. These are standard industry implementations for opaque identifiers and do not conceal malicious commands or instructions.
[EXTERNAL_DOWNLOADS]: The skill references several well-known Node.js SDKs for messaging platforms (Slack, Discord, Telegram, LINE). These are recognized as legitimate tools from well-known services and organizations.

gateway