The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill generates GitHub Action workflows that process untrusted data from GitHub events, creating a potential surface for indirect prompt injection.
Ingestion points: Templates such as ci.yml, e2e.yml, and release.tag.yml ingest data from github.event properties and github.ref_name into the workflow execution context.
Boundary markers: The references/security-hardening.md file provides clear instructions to developers to avoid interpolating untrusted data directly into scripts and to use environment variables as boundaries.
Capability inventory: The generated workflows use run steps for shell execution and request id-token: write for OIDC-based authentication with external services.
Sanitization: The provided templates demonstrate sanitization by passing untrusted values through the env block and double-quoting the resulting shell variables to prevent command injection.
[COMMAND_EXECUTION]: The workflow templates utilize shell execution (run blocks) to perform automation tasks like dependency installation, testing, and building. This is the primary and intended function of the skill.
[EXTERNAL_DOWNLOADS]: The skill references and automates the download of various third-party tools and actions.
Fetches the Turborepo installation script from turborepo.dev (a well-known service).
References numerous official GitHub Actions (e.g., actions/checkout, actions/cache, actions/setup-node) pinned to specific commit SHAs.
Automates the installation of the Bun runtime and Playwright browser binaries from their respective official registries.

github-actions