The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests technology candidates and evaluation criteria from external input to perform experiments. While this creates an attack surface where malicious inputs could attempt to manipulate the testing process, the instructions provide guardrails by requiring explicit success criteria, mandatory evidence recording, and adherence to a "permitted scope" defined at the start of the workflow.
[DYNAMIC_EXECUTION]: The instructions direct the agent to conduct "minimal local experiments" to verify technical feasibility. This involves executing code or interacting with interfaces based on the candidates being evaluated. The risk of arbitrary execution is mitigated by instructions to restrict actions to "permitted interfaces" and "permitted local targets," and by requiring specific approval before modifying any external state.

build-poc