The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on shell command execution through Python's subprocess module to perform its benchmarking tasks. Specifically, scripts/compare.py executes go test -v to verify generated code, while scripts/optimize_loop.py manages git worktree operations to provide isolated environments for evaluation. These are standard operations for a benchmarking framework.
[DYNAMIC_EXECUTION]: The scripts/generate_variant.py and scripts/optimize_loop.py scripts implement an optimization loop that dynamically generates new agent variants. This process involves constructing prompts from existing skill content and executing the Claude CLI (claude -p) to obtain modified versions. This is a functional requirement for the skill's stated purpose of automated research and optimization.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection because it ingests external agent descriptions and benchmark tasks (from JSON files) which are then interpolated into prompts for the scoring agent.
Ingestion points: Processes agents/*.md and JSON files located in references/.
Boundary markers: The skill uses XML-style tags (e.g., <current_content>) to delimit ingested data within prompts.
Capability inventory: Possesses capabilities for file read/write, shell command execution, and sub-agent spawning via the Task tool.
Sanitization: Relies on structural delimiters rather than explicit input filtering. The risk is considered low as it is inherent to the meta-tooling use case and targets a scoring rubric rather than system safety controls.

agent-comparison