The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple shell commands to set up testing environments and verify CLI tools.
Evidence in SKILL.md: Use of git rev-parse, npm ci, npm install, npm run build, npm run typecheck:snippets, and uv venv.
The skill specifically instructs the agent to check shell flags against CLI source or --help output.
[DYNAMIC_EXECUTION]: The skill extracts and executes code blocks (Python and TypeScript) directly from documentation files.
Evidence in SKILL.md: Step 4 instructs writing blocks to .context/verify/ts/<n>.ts and running them via npx tsc, and writing Python blocks to .context/verify/<n>.py and executing them with python <n>.py.
The skill attempts to mitigate risk by limiting execution to the "first model call" unless credentials are provided, but still performs runtime execution of untrusted snippets.
[EXTERNAL_DOWNLOADS]: The skill performs package installations which may fetch code from external registries (NPM, PyPI) based on claims in the documentation.
Evidence in SKILL.md: npm install, uv pip install -e ./strands-py, and reaching out to the GitHub API via gh api repos/strands-agents/<repo>/contents/<path>.
Installations include both local SDKs and external dependencies named in the # pip install ... comments within documentation pages.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process documentation pages which act as untrusted external data sources.
Ingestion points: Files located under site/src/content/docs/ (Step 1).
Boundary markers: Absent. The skill reads the files directly into its processing context without delimiters or sanitization.
Capability inventory: Subprocess calls (npm, uv, python, node, git), file-write (.context/verify), and network operations (gh api).
Sanitization: Absent. There is no escaping or filtering of the code blocks or commands found within the documentation before they are passed to execution tools.

docs-verifier