The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill directs the agent to proactively execute installation commands (npx claude-plugins skills install <namespace>) using parameters retrieved from an external, third-party API (claude-plugins.dev). This allows for the download and execution of arbitrary code from unverified sources.
[DYNAMIC_EXECUTION]: The instructions mandate the creation of new executable skill directories, SKILL.md files, and supporting scripts in the .claude/skills/ directory. The fallback mechanism involves downloading raw content from arbitrary URLs to manually populate these folders, bypassing traditional package manager security checks.
[INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection through its integration with an external search registry.
Ingestion points: Data enters the agent's context from search results at https://claude-plugins.dev/api/skills.
Boundary markers: No explicit delimiters or instructions are used to prevent the agent from being influenced by malicious instructions embedded in the registry's JSON response (e.g., in the 'description' field).
Capability inventory: The agent utilizes npx for package installation and has file system write access to create new skill directories.
Sanitization: The skill relies on 'stars' and 'installs' as quality signals, which are easily spoofed and do not constitute technical sanitization or safety validation of the code being installed.
[COMMAND_EXECUTION]: The skill uses the npx command-line utility to interact with external software packages, which involves spawning subprocesses and running code fetched from the network.

skill-finder