The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the Python subprocess module in several scripts (scripts/run_eval.py, scripts/improve_description.py, and eval-viewer/generate_review.py) to execute local CLI tools like claude and system utilities like lsof. These executions are used to run test cases, optimize skill descriptions, and manage the local evaluation viewer server.
[DYNAMIC_EXECUTION]: The scripts/run_eval.py script dynamically generates temporary skill definition files in the .claude/commands/ directory. These files are used to register and test new skill versions via the claude CLI during the evaluation loop.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves ingesting user-provided prompts and instructions to create or refine AI skills. This creates a surface for indirect prompt injection, as the subagents used for testing and grading could be influenced by malicious instructions embedded in the test data.
Ingestion points: evals/evals.json (test prompts) and user-supplied skill instructions in SKILL.md.
Capability inventory: File writing, subprocess execution, and subagent spawning.
Boundary markers: The scripts use markers like <eval prompt> to delimit input, but lacks explicit "ignore embedded instructions" warnings.
Sanitization: The run_eval.py script utilizes YAML block scalars to help ensure skill descriptions are treated as data rather than control characters.
[EXTERNAL_DOWNLOADS]: The evaluation viewer components (eval-viewer/viewer.html and assets/eval_review.html) reference external resources from well-known services. This includes the SheetJS library from cdn.sheetjs.com and typography from fonts.googleapis.com.

skill-creator