The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to download the @vibetechnologies/chrome-sync package from the npm registry using npx to facilitate browser session synchronization.
[REMOTE_CODE_EXECUTION]: The instruction to use npx -y @vibetechnologies/chrome-sync executes remote code from a third-party source not recognized as a trusted vendor, potentially allowing for the execution of unverified logic.
[COMMAND_EXECUTION]: The skill requires the agent to execute various shell commands, including sourcing environment files (source "$HOME/.hermes/.env"), setting restrictive file permissions (chmod 600), and running authentication checks with tools like gh auth status and git ls-remote.
[CREDENTIALS_UNSAFE]: The skill manages highly sensitive data, including API tokens, SSH keys, and browser session cookies. The use of chrome-sync to push authentication cookies (e.g., for Google accounts) poses a data exposure risk, as it can grant the agent broad access to private services like email and storage if not strictly limited to dedicated profiles.
[PROMPT_INJECTION]: The skill is designed to ingest instructions from messaging gateways (e.g., Telegram), creating an indirect prompt injection surface where untrusted user input can influence sensitive agent actions.
Ingestion points: External messaging gateways as described in the skill metadata.
Boundary markers: The skill uses structured tags (e.g., <role>, <goal>, <inputs>) to define task boundaries and mitigate instruction overlap.
Capability inventory: The agent has capabilities to execute shell commands, write persistent files to ~/.hermes/skills/, and manage system cron jobs.
Sanitization: The workflow includes mandatory read-back verification of written artifacts and state reporting (e.g., SHA-256 digests) to ensure configuration integrity.

prompt-hermes-ai