The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted source code from the local filesystem to perform optimizations.
Ingestion points: The SKILL.md file instructs the agent to use the read_file tool to ingest user-provided code (e.g., read_file("topk_benchmark.cpp")).
Capability inventory: The skill possesses extensive capabilities, including compiling and executing arbitrary code via shell commands (g++, python3, java, rustc, go), writing files to the disk (write_file), and profiling system performance (perf, valgrind).
Boundary markers: The instructions do not specify any delimiters or explicit "ignore embedded instructions" warnings for the agent when processing user-provided code.
Sanitization: There is no evidence of code sanitization or validation before the agent analyzes, modifies, or executes the input code.
[COMMAND_EXECUTION]: The skill relies on executing various system commands to benchmark and profile code performance.
Evidence: SKILL.md provides examples for compiling and running C++, Python, Java, Rust, and Go code using standard CLI tools. It also utilizes profiling tools like perf and valgrind to extract metrics.
[DYNAMIC_EXECUTION]: The skill generates and modifies code at runtime, which is subsequently saved and executed to verify performance gains.
Evidence: The optimization workflow involves writing modified code versions (e.g., topk_benchmark_optimized.cpp) to the filesystem and then executing them via the shell to capture performance metrics.

code-optimization