The Agent Skills Directory

[COMMAND_EXECUTION]: The skill leverages subprocess calls in scripts/run_eval.py, scripts/improve_description.py, and eval-viewer/generate_review.py to interact with the claude CLI for automated testing and to manage local server ports. These operations are scoped to intended functionality and target expected local utilities.
[DYNAMIC_EXECUTION]: The framework dynamically generates and manages files such as markdown instructions, skill templates, and HTML reports. For instance, scripts/run_eval.py temporary registers skills in .claude/commands/ to evaluate triggering accuracy, which is standard procedure for skill development.
[INDIRECT_PROMPT_INJECTION]: The skill manages a surface for indirect prompt injection by processing untrusted outputs and transcripts from other agent runs during the evaluation phase.
Ingestion points: Agent transcripts and generated output files are ingested by subagents as described in agents/grader.md and agents/analyzer.md.
Boundary markers: Instructional prompts use XML-style delimiters to separate system instructions from the content being analyzed.
Capability inventory: The skill maintains the ability to execute claude CLI commands and perform local file operations.
Sanitization: Evaluation results are structured into JSON format, and the process relies on isolated subagent contexts to minimize risk.
[DYNAMIC_CONTEXT_INJECTION]: The skill provides documentation and templates for the !command`` syntax, which is a platform-supported feature for injecting shell output into context at skill load time. The examples provided (e.g., gh pr diff) are contextually appropriate for developer tools and serve as educational material for skill authors.

skill-master