The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from external financial news RSS feeds (e.g., WSJ, CNBC, Reuters) which is subsequently analyzed by an LLM to generate summaries. This establishes a surface for indirect prompt injection where adversarial content in a headline could attempt to manipulate the agent.
Ingestion points: RSS feeds are aggregated in scripts/fetch_news.py and passed to summarization logic.
Boundary markers: The scripts/summarize.py file includes a HARDENED_SYSTEM_PROMPT designed to instruct the LLM to ignore any instructions or commands embedded in the processed data.
Capability inventory: The skill can send messages to WhatsApp or Telegram using the openclaw message tool.
Sanitization: The skill relies on specific system prompt instructions and LLM-native guardrails to sanitize processed headlines.
[COMMAND_EXECUTION]: The skill utilizes subprocess.run to orchestrate various Python scripts and external CLI tools.
Evidence includes briefing.py calling summarize.py, summarize.py calling the openclaw agent and gemini CLI, and fetch_news.py executing openbb-quote. These calls are used for standard functional orchestration rather than processing raw shell commands.

finance-news