The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill uses npx --yes impeccable@latest to install and update the impeccable tool directly from the NPM registry. While NPM is a standard service, the tool itself is not from a pre-defined trusted source.
[COMMAND_EXECUTION]: The instructions involve executing various CLI tools, including npx and claude. The skill follows security best practices by using standard input redirection from temporary files (e.g., claude ... < /path/to/handoff.txt) rather than interpolating text into shell commands, which prevents command injection vulnerabilities.
[INDIRECT_PROMPT_INJECTION]: The skill has a vulnerability surface where instructions embedded in external project data could influence the agent's behavior.
Ingestion points: The skill reads PRODUCT.md, design documents, plans, and the current working-tree state to build handoff prompts.
Boundary markers: The instructions do not specify the use of delimiters or 'ignore' instructions for the content being interpolated into the delegate prompts.
Capability inventory: The skill executes shell commands, writes files (e.g., PRODUCT.md), and launches a delegate agent with the ability to implement code and potentially access the network via the claude CLI.
Sanitization: The skill avoids shell interpolation by using file redirection for prompts, but it does not specify sanitization or validation of the text content extracted from design documents and project files.
[SAFE]: The skill includes specific defensive instructions, such as verifying that cleanup targets are not symlinks and explicitly prohibiting the use of --dangerously-skip-permissions or bypassPermissions flags.

delegate-ui-to-claude