The Agent Skills Directory

[COMMAND_EXECUTION]: The skill workflow begins by instructing the agent to execute npx skills add christopherlhammer11-ai/tool-use-guardian. This command uses the Node.js package runner to download and execute code or configuration from a third-party, unverified source.
[EXTERNAL_DOWNLOADS]: The skill references and attempts to install components from multiple unverified external repositories and domains, including github.com/christopherlhammer11-ai/tool-use-guardian and github.com/sickn33/antigravity-awesome-skills. It also references non-whitelisted domains such as genesis-node-api.vercel.app and genesis-marketplace.vercel.app for skill management.
[PROMPT_INJECTION]: The skill is designed to intercept, 'fix', and 'auto-coerce' tool outputs and malformed JSON. This creates a surface for indirect prompt injection (Category 8) where an attacker-controlled tool output could bypass security boundaries during the 'fixing' process.
Ingestion points: Processes truncated JSON, API timeouts, and error messages from external tools as described in the 'Failure Classification' section.
Boundary markers: No explicit markers are defined to prevent the agent from obeying instructions embedded within the tool data it is attempting to fix.
Capability inventory: The skill possesses command execution capabilities (npx) and manages multi-step tool chains.
Sanitization: The workflow describes 'auto-coercion' and 'auto-recovery' from malformed data without specifying any validation or sanitization protocols to ensure the integrity of the recovered content.

tool-use-guardian