The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from the repository (AGENTS.md, .AGENTS.md, etc.) and evaluates it against quality criteria. Malicious instructions within these files could attempt to hijack the agent's logic during assessment.
Ingestion points: Reads file content via Read, Glob, and Grep in SKILL.md (Phase 1 and 2).
Boundary markers: None present to distinguish between data and instructions.
Capability inventory: Access to bash for file discovery and Edit for file modification.
Sanitization: No evidence of sanitization or instruction-ignoring delimiters when processing the file content.
[METADATA_POISONING]: All reference files (references/quality-criteria.md, references/templates.md, and references/update-guidelines.md) refer to CLAUDE.md while the primary skill instructions and tools target AGENTS.md. This inconsistency is likely a documentation error but could lead to schema confusion or misinterpretation of instructions by the agent.

Codex-md-improver