The Agent Skills Directory

[PROMPT_INJECTION]: The skill requests high autonomy, explicitly instructing the agent to "do not pause to ask user feedback" once the experiment loop starts. This reduces human oversight over the modification and execution of code.
[PROMPT_INJECTION]: The skill is susceptible to Indirect Prompt Injection. It ingests untrusted kernel source code from the src/tilegym/ directory and executes it via pytest to verify correctness.
Ingestion points: Reads kernel files under src/tilegym/suites/ or src/tilegym/ops/ (specified in SKILL.md).
Boundary markers: Absent. No instructions are provided to delimit or ignore embedded instructions within the source files.
Capability inventory: Includes git, python -m pytest (code execution), pip install, and tileiras (compiler).
Sanitization: Absent. The agent is designed to parse, modify, and execute the content of the files directly.
[COMMAND_EXECUTION]: The skill implements an iterative experimentation loop that modifies Python source files and executes them. This includes a "Creative Optimization Plan" (Optimization L) where the agent is encouraged to generate and apply custom, non-templated logic to the codebase.
[EXTERNAL_DOWNLOADS]: The ir-dump-guide.md documentation instructs the agent to install the cuda-tile package via pip. While this appears to be a legitimate vendor-associated package, runtime installation of external dependencies is a privileged operation that could be abused.
[REMOTE_CODE_EXECUTION]: The skill involves a complex toolchain including tileiras for bytecode compilation and cuobjdump for SASS analysis. The agent executes these tools on generated artifacts, creating a path for code execution based on processed data.

improve-cutile-kernel-perf