The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Python's subprocess module to perform various tasks:
scripts/run_eval.py uses subprocess.Popen to execute claude -p commands for running skill evaluations.
eval-viewer/generate_review.py uses subprocess.run to call lsof (to manage port conflicts) and taskkill (on Windows) to reap processes.
These actions are consistent with the skill's purpose as a development and testing harness.
[EXTERNAL_DOWNLOADS]: The evaluation review viewer (eval-viewer/viewer.html) references the SheetJS library (xlsx.full.min.js) via a CDN (cdn.sheetjs.com). This is a well-known library for processing spreadsheet data and is used here for displaying eval outputs.
[INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection as it is designed to ingest and execute user-provided test prompts (evals/evals.json) through the run_eval.py script.
Ingestion points: The --eval-set parameter in scripts/run_eval.py and scripts/run_loop.py loads external JSON files containing prompts.
Boundary markers: The scripts pass the prompts directly to the CLI without specific sanitization beyond list-based subprocess arguments.
Capability inventory: The skill can execute CLI commands, write files, and perform network requests (via the anthropic client in scripts/improve_description.py).
Sanitization: None observed for the prompt content itself; the skill relies on the underlying LLM's safety guardrails.
[DYNAMIC_CONTEXT_INJECTION]: The documentation (SKILL.md and references/workflows.md) explicitly describes and encourages the use of !command syntax for fetching live data during skill execution. This is an intended feature of the target platform (Claude Code) being documented for developers.
[PRIVILEGE_ESCALATION]: The scripts/init_skill.py script uses chmod(0o755) to ensure generated helper scripts are executable by the user. This is standard behavior for a scaffolding tool.

skill-creator