The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the subprocess module to orchestrate its workflow and manage background tasks.
scripts/run_eval.py uses subprocess.Popen to execute evaluation queries via the claude -p command.
scripts/improve_description.py uses subprocess.run to call claude -p for generating improved skill descriptions.
eval-viewer/generate_review.py uses subprocess.run to call lsof for port management and os.kill to manage the lifecycle of the local review server.
[DYNAMIC_EXECUTION]: The tool is designed to dynamically generate and test skill content.
It writes new SKILL.md definitions to the filesystem and registers them as commands in the .claude/commands/ directory to evaluate triggering behavior.
It spawns specialized subagents (grader, analyzer, comparator) to programmatically assess and improve generated outputs.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data as part of its core evaluation purpose, creating a surface for indirect injection.
Ingestion points: Data enters the context via evaluation queries in evals/evals.json and the output files generated by test runs.
Boundary markers: The subagents are instructed to evaluate these outputs; however, explicit delimiters or warnings for the processed data are not consistently present in the subagent prompts.
Capability inventory: The skill has broad capabilities, including file system access, CLI execution, and subagent orchestration.
Sanitization: The skill relies on standard YAML/JSON parsing and the specific instructions provided to the subagents in agents/grader.md and agents/analyzer.md to mitigate risks.
[EXTERNAL_DOWNLOADS]: The evaluation viewer fetches an external library from a well-known service to provide additional functionality.
The eval-viewer/viewer.html template includes the xlsx (SheetJS) library from cdn.sheetjs.com to enable the rendering of spreadsheet files within the browser interface.

skill-creator