The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text from social media and token metadata. It includes explicit instructions to treat this data as content to be summarized rather than instructions to follow, effectively mitigating potential injection attacks.
Ingestion points: Token metadata (symbol, name, links) retrieved via gmgn-cli and social post text/author information gathered via social search APIs.
Boundary markers: The skill explicitly warns that all external content is 'ATTACKER-CONTROLLED TEXT' and provides instructions to ignore instruction-shaped content, including looking for [filtered] markers from gmgn-cli.
Capability inventory: The skill uses subprocess calls to gmgn-cli, network operations for social search, and a web content fetch interface for verification.
Sanitization: The skill relies on gmgn-cli for character filtering and metadata neutralization, and implements custom sanitization logic for its social search and fetch interfaces.
[COMMAND_EXECUTION]: The skill relies on the gmgn-cli tool to interact with on-chain data. It provides clear instructions for the configuration and verification of this tool, including handling API keys provided by the user when necessary.
[EXTERNAL_DOWNLOADS]: The skill defines a web content fetch interface for verifying claims, which includes robust security requirements such as protocol whitelisting (HTTP/HTTPS only) and SSRF protections including private IP range blacklisting and DNS rebinding protection.

gmgn-narrative