The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The init command performs an automatic check and update of AI coding skills from a remote repository (GitHub). The browser ensure command downloads a specific, pinned version of the Chrome browser to the local machine to ensure rendering consistency.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources via the capture command, which processes live URLs into editable projects, and the validate command, which executes project code in headless Chrome.
Ingestion points: references/init-and-scaffold.md (via capture <URL>) and references/lint-validate-inspect.md (via validate and inspect loading compositions).
Boundary markers: No specific delimiters or "ignore embedded instructions" warnings are documented for the ingestion of external web content.
Capability inventory: The CLI has access to shell execution via npx, file system writes, and AWS cloud deployment capabilities.
Sanitization: The CLI implements path redaction (converting $HOME to a literal string) in JSON output to prevent leaking local environment details in logs shared with agents.
[DATA_EXFILTRATION]: The telemetry feature fingerprints the execution environment to distinguish sandboxes (e.g., gvisor, firecracker) and specific AI agent environments (e.g., claude_code, cursor, replit) based on the presence of vendor-specific environment variables. While the skill claims values are not read, the presence of these markers is transmitted externally.
[DYNAMIC_EXECUTION]: The lambda deploy workflow involves building a deployment package (handler.zip) at runtime. Furthermore, the init command's behavior of automatically pulling and updating 'latest skills' from GitHub represents a dynamic update mechanism for the agent's own operating instructions.

hyperframes-cli