The Agent Skills Directory

[PROMPT_INJECTION]: As a generative tool that creates skills based on user input, the skill inherently possesses an attack surface for indirect prompt injection. A malicious user could provide input designed to create a harmful skill.
Ingestion points: User-provided descriptions and goals for new skills or agents in SKILL.md.
Boundary markers: The skill does not currently implement specific delimiters or "ignore" instructions within its generated templates to isolate user-provided logic.
Capability inventory: The skill is designed to perform file system operations (read, mkdir, write) and execute platform-specific commands (openclaw gateway restart, python3) via the agent's available tools.
Sanitization: The skill relies on the underlying LLM's general safety guardrails and does not perform explicit sanitization of the user's creative input.
[EXTERNAL_DOWNLOADS]: The skill includes an example for a weather bot that uses curl to fetch data from wttr.in. This is a well-known service used for legitimate demonstration of network capabilities and does not involve downloading executable code.
[SAFE]: The main instructions guide the agent to execute local validation scripts (e.g., quick_validate.py) located within the system's openclaw installation using standard platform commands and paths.

agent-forge