The Agent Skills Directory

[COMMAND_EXECUTION]: Multiple Python scripts (scripts/run_eval.py, scripts/improve_description.py, scripts/run_loop.py, eval-viewer/generate_review.py) use the subprocess module to execute shell commands. This includes running the claude CLI for evaluations and improvements, as well as system utilities like lsof for process management.
[EXTERNAL_DOWNLOADS]: The evaluation viewer (eval-viewer/viewer.html) loads the SheetJS library from a public CDN (cdn.sheetjs.com) to enable spreadsheet rendering for the user.
[COMMAND_EXECUTION]: The render-graphs.js utility executes the system dot command (Graphviz) to generate SVG diagrams from DOT files.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input in the form of user-defined evaluation queries and conversation history. This content is interpolated into prompts and passed to the claude CLI, presenting a surface for indirect prompt injection.
Ingestion points: Evaluation queries from user-provided eval_set.json and workflow context extracted from conversation history.
Boundary markers: Employs XML-style tags (e.g., <current_description>, <skill_content>) to structure prompts and separate instruction from data.
Capability inventory: Extensive local command execution (subprocess), file system writes (.opencode/commands/), and workspace management.
Sanitization: Uses YAML structural escaping for description fields in scripts/run_eval.py, but evaluation queries are generally passed as-is to the underlying model.

creating-skills