The Agent Skills Directory

[METADATA_POISONING]: The skill's YAML frontmatter contains a createdBy field referencing a non-existent version of a language model ('gpt-5.6-sol'). This metadata is misleading and does not accurately reflect the origin or creation process of the skill content.
[INDIRECT_PROMPT_INJECTION]: The inspect-project.py script parses external project files such as Gradle scripts and Kotlin sources. This creates a surface where a malicious project could embed instructions within configuration strings (e.g., benchmark target names) that the agent might follow when interpreting the script's output.
Ingestion points: The script reads file contents from the project directory using Path.rglob and read_text in the inspect function.
Boundary markers: The script returns data in structured JSON format, but the skill does not explicitly instruct the agent to ignore instructions that may be embedded within this data.
Capability inventory: The agent uses the script's output to verify toolchain versions and guide users through benchmark setup, as specified in SKILL.md.
Sanitization: The script extracts values using regular expressions without sanitizing the resulting strings for prompt injection patterns.
[COMMAND_EXECUTION]: The skill instructs the agent to run a bundled Python script (scripts/inspect-project.py) to gather project metadata. The script is restricted to local read operations on the project file system.
[EXTERNAL_DOWNLOADS]: The skill provides links to the official kotlinx-benchmark repository and documentation on GitHub (github.com/Kotlin/kotlinx-benchmark). These are well-known and trusted resources for the tool's intended use.

kotlinx-benchmark