The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources, such as knowledge files (KNOWLEDGE.jsonl), project documentation, and session history, to generate project-wide rules. These rules are stored in .claude/rules/ and are automatically loaded by the AI platform in subsequent sessions, potentially allowing malicious content in processed documents to influence future agent behavior.
Ingestion points: The skill reads content from user-specified file paths and session learnings as described in the SKILL.md instructions and README.md examples.
Boundary markers: While the skill uses specific markers (e.g., t:"❌" and t:"✅") to structure the data, it lacks robust boundary protections to prevent instructions within the ingested data from being executed as rules.
Capability inventory: The skill possesses the capability to Write and Edit project files, specifically updating markdown files that dictate agent behavior across sessions.
Sanitization: No explicit sanitization or filtering is performed on the content extracted from external documents before it is written to the persistent rule files.
[COMMAND_EXECUTION]: The skill utilizes a local bash script (scripts/rules.sh) to perform file management, validation, and status reporting. The script is invoked with parameters derived from the user prompt or agent reasoning.
The script incorporates safety checks, such as validating the YAML frontmatter of generated files and refusing to apply specialized rules to the entire repository root to prevent drift.

rules