The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from external platforms (X/Twitter, Reddit, GitHub, Amazon, and YouTube), creating a potential attack surface where malicious instructions could be embedded in the fetched data.
Ingestion points: The agent fetches full-page extractions, snippets, and transcripts from arbitrary web URLs and social media platforms via the mcp__glim__* tools.
Boundary markers: The skill provides explicit instructions to the agent to treat all fetched content as untrusted data and specifically commands it to ignore and discard sources containing phrases like "ignore previous instructions."
Capability inventory: The skill possesses the ability to execute local scripts (scripts/mcp-oauth-refresh.sh), call a notification utility (./notify), and interact with external APIs.
Sanitization: The skill relies on instructional guardrails to identify, discard, and log malicious content encountered in external sources rather than programmatically sanitizing the input.
[COMMAND_EXECUTION]: The skill executes a local script, scripts/mcp-oauth-refresh.sh, to manage the lifecycle of OAuth tokens (MCP_GLIM_TOKEN and MCP_GLIM_OAUTH) required for the glim.sh service.

glim-mcp