The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The installation process involves downloading a ZIP archive from a third-party, untrusted GitHub repository at https://github.com/oalanicolas/claude-design-premium/archive/refs/heads/main.zip.
[REMOTE_CODE_EXECUTION]: The skill requires the execution of multiple Node.js scripts (bootstrap-harness.mjs, context-signals.mjs, test-builder-bootstrap.mjs) included in the downloaded archive to perform setup and validation tasks. Executing downloaded code from an unverified source is a significant security risk.
[COMMAND_EXECUTION]: The skill provides instructions for the agent to execute shell commands such as curl, unzip, and node to install and run the harness.
[INDIRECT_PROMPT_INJECTION]: The skill processes external files that could contain malicious instructions designed to manipulate agent behavior.
Ingestion points: Files like DESIGN.md, _ds_manifest.json, and BOUND_DS.json are read and interpreted to set design constraints and component rules.
Boundary markers: The skill lacks explicit boundary markers or "ignore instructions" warnings when processing the contents of these configuration files.
Capability inventory: The skill scripts possess file system access (read/write) and shell execution capabilities, which could be abused if malicious data is ingested.
Sanitization: There is no evidence of sanitization or validation for the natural language rules or component definitions read from external files.

claude-design-premium-harness