The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes local text files (agent rules, project memory, and user notes) to extract behavioral patterns and personality traits.
Ingestion points: scripts/extract_rules.py reads data from ~/.claude/CLAUDE.md, project-level CLAUDE.md/AGENTS.md files, and project memory markdown files.
Boundary markers: The generated templates (templates/SKILL.template.md) lack explicit prompt delimiters for interpolated rules, though the skill instructions require the user to manually confirm or edit the extracted rules before the skill is built.
Capability inventory: The skill scripts have local file system read/write access to user-specified directories and the agent's skills folder.
Sanitization: The skill implements extensive regex-based redaction in scripts/extract_rules.py for URLs, emails, IPs, phone numbers, and names. It also includes a validation script (scripts/check_clone.py) that scans generated files for leaked PII before completion.
[DYNAMIC_EXECUTION]: The provided Python scripts use dynamic module loading to share patterns and validation logic between components.
Evidence: scripts/build_clone.py and scripts/check_clone.py utilize importlib.util.spec_from_file_location to load one another as modules at runtime.
Mitigation: The paths for dynamic loading are strictly resolved using Path(__file__).resolve().parent, which ensures that only local scripts within the skill's own directory can be loaded, preventing the execution of arbitrary external code.

clone-yourself