The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes Python's subprocess and os modules to manage the development lifecycle, including executing the claude CLI for evaluations, controlling a local web server for the results viewer, and automating packaging and benchmarking tasks.
[EXTERNAL_DOWNLOADS]: The eval-viewer/viewer.html component fetches the SheetJS library from cdn.sheetjs.com. This is a well-known service used to provide spreadsheet rendering capabilities within the review interface.
[DYNAMIC_EXECUTION]: The grader subagent instructions (agents/grader.md) suggest that the agent can generate and execute scripts at runtime to programmatically verify test assertions. This is an intended feature of the testing harness.
[INDIRECT_PROMPT_INJECTION]: The skill has an indirect injection surface as it processes external data from evaluation files and user feedback to inform skill improvements.
Ingestion points: Reads evals/evals.json, feedback.json, and execution transcripts (referenced in SKILL.md and agents/grader.md).
Boundary markers: Absent in the prompts generated for the skill optimizer.
Capability inventory: Includes subprocess execution, file writes, and network access via platform tools (documented in scripts/run_eval.py and scripts/improve_description.py).
Sanitization: Content from feedback and evaluation outputs is interpolated into optimizer prompts without specific escaping or structural validation.

skill-creator