The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the Bash tool to perform repository maintenance, specifically using git checkout to revert changes if the improvement loop does not result in a better score.
[PROMPT_INJECTION]: The skill processes external files (SKILL.md files) and generates content based on their instructions, which presents a surface for indirect prompt injection.
Ingestion points: Reads arbitrary skill files from the skills/ directory in Phase 3.
Boundary markers: The skill explicitly prompts the user for permission ("May I write this improved version...?") in Phase 4 before modifying the file system.
Capability inventory: Includes Read, Write, and Bash tools, allowing for file modification and shell command execution.
Sanitization: No technical sanitization of the input file content is performed; the skill relies on the underlying LLM's reasoning and the final user approval step.

skill-improve