The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text from files which could contain embedded instructions intended to influence agent behavior.
Ingestion points: Reads local files using view_file as defined in agents/openai.yaml.
Boundary markers: Implements a 'Structural AST Shield' protocol to tokenize and isolate code blocks, tables, and frontmatter from prose processing, as detailed in references/ast-shield-and-cadence.md.
Capability inventory: Includes write_to_file and replace_file_content for modifying local documents.
Sanitization: Enforces strict factual fidelity, non-fabrication of data, and byte-parity verification for technical elements to prevent unauthorized content generation.

humanize