The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external GitHub comments to build a behavioral model, which is a potential surface for indirect prompt injection if an attacker submits malicious comments to influence the reviewer clone.
Ingestion points: Interaction data including comments are fetched via the GitHub CLI and stored in scratch/comments/ for the trainer to process.
Boundary markers: The workflow does not explicitly mention the use of delimiters or warnings to ignore instructions embedded within the human comments being analyzed.
Capability inventory: The generated 'Clone' agent has the ability to post reviews and comments to GitHub using gh pr review and gh api.
Sanitization: There is no specified logic for sanitizing or filtering instructions out of the comment text during the learning phase.
[DYNAMIC_EXECUTION]: The skill's primary function involves the dynamic generation of a new AI agent ('the Clone') by writing a new SKILL.md, MODEL.md, and VOICE.md to the user's local skill directory.
[COMMAND_EXECUTION]: The skill uses subprocess.run to execute local git and gh commands for repository analysis and data collection, as seen in scripts/collect-review-activity.py and the workflow.md instructions.

melech-cr-clone-trainer