The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves downloading and analyzing external agent skills from GitHub and third-party marketplaces. This untrusted content is directly interpolated into LLM prompts for assessment and security auditing, creating a vulnerability to indirect prompt injection where a malicious skill could attempt to manipulate the analysis results.
Ingestion points: scripts/assess.ts and scripts/audit.ts read SKILL.md and script files from downloaded repositories using Bun.file().text().
Boundary markers: The prompt templates wrap external content in markdown code blocks, which can be subverted by adversarial input.
Capability inventory: The skill possesses the ability to execute shell commands (gh, npx, find, grep) and write files to the filesystem.
Sanitization: There is no evidence of robust sanitization or escaping of the external content before it is processed by the agent.
[COMMAND_EXECUTION]: The system relies on executing several external command-line tools to manage skills.
scripts/pipeline.ts and scripts/search.ts invoke the GitHub CLI (gh) to search for and clone remote repositories.
scripts/skills-cli.ts uses npx skills to interact with the Vercel Labs skill ecosystem for discovery and installation.
scripts/assess.ts and scripts/audit.ts use standard Unix utilities like find, grep, and wc to analyze the local directory structure and file contents of downloaded skills.
[EXTERNAL_DOWNLOADS]: The skill fetches metadata and source code from external network locations.
scripts/search.ts retrieves trending skill data from the skills.sh marketplace via HTTPS.
scripts/pipeline.ts clones external code repositories from GitHub to the local machine.
scripts/skills-cli.ts triggers the download and execution of the skills utility from the NPM registry via npx.

skill-management