The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill depends on numerous Python scripts located in scripts/skill_eval/ (such as run_eval.py, run_loop.py, and quick_validate.py) and a script at ~/.claude/scripts/learning-db.py that are not included in the provided files. Because these scripts are not available for audit, their safety and actual behavior remain unverifiable. The use of a path in the user's home directory for core logic is non-standard.
[COMMAND_EXECUTION]: The skill instructions direct the agent to execute various shell commands, including python3 module calls and the claude CLI, to run trigger tests, benchmarks, and optimization loops.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and evaluate external, potentially adversarial content from evals.json and the outputs of other skills, creating an attack surface for indirect prompt injection.
Ingestion points: The system reads evals.json for test prompts and processes execution transcripts and file outputs from the skills being benchmarked.
Boundary markers: The provided sub-agent instructions (grader.md, comparator.md) do not include explicit delimiters or instructions to ignore embedded commands when processing these outputs.
Capability inventory: The skill has access to Bash, Write, Read, Grep, Glob, and Agent tools, which could be abused if the agent is influenced by malicious instructions in the data it evaluates.
Sanitization: There is no description of sanitization, filtering, or validation of the input prompts or the skill outputs before they are processed by the evaluation logic.
[DYNAMIC_EXECUTION]: The skill implements automated self-improvement and optimization processes (run_loop.py, self-improve-loop.md) that generate new variants of SKILL.md instruction files and promote them to replace the original definitions. The automated generation and promotion of executable instructions constitutes dynamic execution of generated content.

skill-eval