The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted codebase data including source code, tests, and documentation, creating an indirect prompt injection surface. Malicious instructions embedded in analyzed files could potentially attempt to influence the agent's architectural recommendations.\n
Ingestion points: The skill is instructed to read repo instructions, workflow configurations, caller code, and specifications in SKILL.md and references/assessment.md.\n
Boundary markers: There are no instructions to use delimiters or ignore-behavior warnings when ingesting external code content.\n
Capability inventory: The skill is restricted to advisory work and is explicitly barred from implementing code or modifying the repository state, which significantly reduces the risk of malicious action.\n
Sanitization: No methods for escaping or validating external codebase content are specified.\n- [EXTERNAL_DOWNLOADS]: The skill's reference documentation includes links to external engineering guidelines hosted on GitHub for attribution.\n
Evidence: Links to the mattpocock/skills repository appear in references/assessment.md and references/codebase-design.md.\n
Context: These are informational source attributions and do not represent runtime code downloads or execution.

ziw-architecture