The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fannouts research queries to numerous external platforms, including Reddit, X/Twitter, YouTube, GitHub, Hacker News, Polymarket, and various other social networks. Most traffic is routed through the surf.cascade.fyi API, while some sources use direct HTTP requests to public JSON endpoints for free-tier access.
[COMMAND_EXECUTION]: Subprocess execution is used appropriately for development and metadata retrieval. lib/github.py attempts to run the gh (GitHub CLI) tool to fetch authentication tokens. Testing and verification scripts (verify_v3.py, evaluate.py) use subprocess to run local Python scripts and git commands for worktree management during evaluations.
[CREDENTIALS_UNSAFE]: The skill manages sensitive API keys (SURF_API_KEY, GITHUB_TOKEN) using environment variables and configuration files. It includes a security helper in lib/env.py that checks the file system permissions of the .env configuration file and warns the user if it is world-readable, recommending chmod 600 for protection.
[PROMPT_INJECTION]: The skill is specifically hardened against indirect prompt injection (attacks embedded in the scraped research data). Both the query planner and the reranking logic use a mandatory <untrusted_content> fence in LLM prompts, accompanied by clear instructions to the agent to treat the enclosed content strictly as data and never as executable instructions.

last30days-surf