The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted source code from external directories (Ingestion points: SKILL.md, hunt.md, map.md, trace.md, study.md). It proactively mitigates the risk of malicious instructions within that code using an 'untrusted-content envelope' (Boundary markers: Present in all modules) that directs the agent to treat source code as data only. The skill possesses the capability to read files and execute specific platform-specific scripts (Capability inventory: libexec/raptor-* utilities). Explicit instructions to flag and ignore instruction-shaped text provide additional security (Sanitization: Present).
[COMMAND_EXECUTION]: The skill relies on various 'raptor-*' utility scripts. These operations are managed via a platform-level allowlist (noted in SKILL.md as '.claude/settings.json') which restricts the agent to pre-approved commands and requires operator permission for others.
[SAFE]: Deterministic triggers for 'ignore previous instructions' are false positives; these strings are used within training examples for the agent to identify and ignore prompt injection attempts in analyzed code.

code-understanding