The Agent Skills Directory

[PROMPT_INJECTION]: Indirect prompt injection vulnerability surface. The skill ingests untrusted data from external sources (NPM, PyPI, GitHub, and Web searches) to guide implementation decisions.
Ingestion points: Results from package registry metadata and general web searches.
Boundary markers: Absent.
Capability inventory: Includes command execution via package managers and file system writes for implementation.
Sanitization: Not specified; relies on the agent's evaluation of search results.
[COMMAND_EXECUTION]: Suggests executing shell commands for auditing local code and managing dependencies.
Evidence: Instructions to use rg (ripgrep) for searching the repository and npm install or pip install for library adoption.
Context: These are standard operations for a coding agent but represent a capability that could be misused.
[EXTERNAL_DOWNLOADS]: Encourages the installation of third-party packages based on search findings.
Evidence: Explicit instructions to adopt and install packages like ajv-cli or textlint-rule-no-dead-link.
Context: While directed at reputable registries, the dynamic nature of selection presents a supply-chain risk.

search-first