The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions define a process that ingests user-provided optimization goals and variant hypotheses to be executed via system tools. This creates an attack surface where untrusted data could influence generated commands.
Ingestion points: Performance targets and variant definitions (Hypothesis/Command) described in the 'Variant Table' and 'Recursive Search' sections of SKILL.md.
Boundary markers: The skill explicitly instructs the agent to use a 'correctness gate' and 'reject variants that fail correctness, safety, or reproducibility', providing some mitigation against accidental execution of malicious code.
Capability inventory: The skill utilizes Bash, Write, Edit, and Grep tools, which allow for code modification and execution.
Sanitization: The skill relies on the agent's evaluative reasoning to reject unsafe variants but lacks technical sanitization or isolation protocols for the variant commands.
[DYNAMIC_EXECUTION]: The primary purpose of the skill is to generate, promote, and execute code variants at runtime to measure performance deltas.
The instructions guide the agent to 'Generate variants that test one hypothesis each' and 'Run variants', which involves dynamic creation and execution of scripts or commands through the Bash tool.

benchmark-optimization-loop