The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides a mechanism to update and manage a project-level rules file (.claude/RULES.md) which directly influences agent behavior and decision-making.
Ingestion points: Untrusted user input is ingested through rule creation requests (e.g., "new rule", "add a rule") as described in SKILL.md.
Boundary markers: The skill implements structural boundaries by requiring a three-part markdown format (Title, Directive, Rationale) and validates entries against a JSON schema (assets/schemas/rule-entry.schema.json).
Capability inventory: The skill reads (cat) and generates content for behavioral directives stored in a persistent file system location. These rules are explicitly described as binding for all agents in the repository.
Sanitization: The included validation script (scripts/validate-rules.sh) utilizes yq with strenv() and jq with raw string input (-R -s) to safely handle and escape user-provided strings during the rule generation process, mitigating command or tool-specific injection during the write phase.

rules-management