The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run across several utility scripts (scripts/autoevolve.py, scripts/run-eval.py, scripts/grade.py, scripts/test-triggers.py) to interact with the platform's CLI (claude or claude-code). This is intended for running tests, grading assertions, and performing trigger analysis.
[COMMAND_EXECUTION]: The SKILL.md instructions guide the agent to execute local scaffolding and validation commands using uv run. These operations are limited to the skill's local directory and standard configuration paths.
[DATA_EXFILTRATION]: The scripts read and write files within local skill directories (~/.claude/skills or .claude/skills) to manage instruction files, reference documents, and evaluation logs. While this involves filesystem access, it is restricted to the skill development environment.
[PROMPT_INJECTION]: The scripts/autoevolve.py script implements an autonomous evolution loop that ingests LLM-generated proposals to modify SKILL.md. This represents an indirect prompt injection surface where the agent can rewrite its own instructions. However, this is a core functional feature for automated skill optimization and is intended for use in a development context.
Ingestion points: The script scripts/autoevolve.py reads proposals for SKILL.md directly from LLM output.
Boundary markers: The generated instructions are written to SKILL.md without explicit boundary markers or sanitization of embedded commands.
Capability inventory: The skill possesses the capability to write files and execute shell commands via the claude CLI and subprocess.run calls in its utility scripts.
Sanitization: No specific sanitization or filtering is applied to the instructions generated during the evolution loop beyond basic JSON parsing of the streaming output.

skill-creator