The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches research data from well-known and reputable public services including Reddit, Hacker News (Algolia API), and established news publishers (e.g., NYT, Wired). These operations are restricted to unauthenticated public APIs and standard web search tools.
[COMMAND_EXECUTION]: Utilizes local Python scripts (citation_tracker.py, time_window_calculator.py, topic_slug_generator.py) for utility tasks. These scripts are limited to the Python standard library and perform benign operations like timestamp calculation, filesystem-safe slug generation, and local session logging.
[SAFE]: Incorporates 'Agent Integrity Rules' that enforce strict source discipline. It requires the agent to cite only URLs returned by tool calls in the current session, effectively mitigating the risk of citation hallucination and training data leakage.
[SAFE]: Implements structured 'Grill-Me Intake' questions to ensure topic specificity and narrow search parameters before any network operations occur, minimizing unnecessary resource usage and improving output relevance.
[PROMPT_INJECTION]: The skill processes untrusted external data from the web (indirect prompt injection surface). However, it includes robust mitigations: (1) Ingestion points: Reddit, HN, and WebFetch tool results in SKILL.md. (2) Boundary markers: Explicit instructions in the Agent Integrity Rules to cite only tool-call results and label training knowledge. (3) Capability inventory: Access to local utility scripts and standard web search/fetch tools. (4) Sanitization: Logic in the Synthesis section requires the agent to analyze and paraphrase findings into a structured briefing format.

pulse