The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes several local Python scripts and shell commands (e.g., generate_review.py, aggregate_benchmark, run_loop.py, nohup, kill) to manage the lifecycle of skill creation and evaluation.
[COMMAND_EXECUTION]: It accesses the agent's internal trajectory data at ~/.claude/projects/ to generate test cases from previous sessions.
[PROMPT_INJECTION]: The skill instructs the agent to intentionally influence model behavior by creating 'pushy' skill descriptions to increase triggering frequency.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting and processing untrusted data from trajectory logs and user feedback to generate instructions.
Ingestion points: Reads failed_trajectories.jsonl, trajectory_samples.jsonl, and feedback.json to inform skill modifications.
Boundary markers: Includes a 'Principle of Lack of Surprise' and guidelines against creating malicious content to mitigate risks.
Capability inventory: The skill has significant capabilities including file system access, shell command execution, and subagent orchestration.
Sanitization: No explicit technical sanitization or validation of the ingested trajectory or feedback data is mentioned.

skill-creator