The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to "Interview the repository" by reading source code and documentation. This creates an attack surface where a malicious repository could contain instructions designed to influence the agent's behavior during the harness construction process.
Ingestion points: The entire repository being analyzed (Step 1).
Capability inventory: The agent is directed to define and execute shell commands for launching, driving, and cleaning up the product.
Boundary markers: None specified for the ingested repository content.
Sanitization: The skill advises redacting secrets and personal data from evidence, but does not specify sanitization for the instructions ingested from the repo.
[COMMAND_EXECUTION]: The skill's primary purpose is to define and execute shell commands to build and verify software products. It mitigates risk by requiring "unique run IDs", "isolated resources", and "resource ownership records" to ensure the agent only interacts with and cleans up processes it started.
[DYNAMIC_EXECUTION]: The skill guides the agent in generating a new verifying-<product> skill and associated helper code at runtime. It includes a "Doctor" phase to verify the identity and safety of the generated harness before execution.

building-verification