The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill has a functional surface for indirect prompt injection because it is designed to read and analyze untrusted content from external code repositories (READMEs, comments, and documentation). However, the skill explicitly mitigates this risk through rule ASI01 (Goal Hijack), which instructs the agent to treat all repository content as passive reference material and to disregard any instructions embedded within it.
Ingestion points: read_file, search_code, and search_repo tools.
Boundary markers: The skill includes explicit instructions to treat summary outputs as untrusted inputs for downstream agents and to separate confirmed facts from assumptions.
Capability inventory: The skill possesses execute_command, write_file, and run_build capabilities, which are necessary for service exploration but are restricted by the core rules requiring verification of all AI-suggested locations.
Sanitization: The skill mandates stripping internal workflow terminology and PII/secrets before generating the navigation summary.
[COMMAND_EXECUTION]: The skill frontmatter allows for the execution of commands (execute_command, run_build, run_dev_server) to assist in mapping service behavior. These capabilities are used within the context of developer orientation and are balanced by instructions to verify runtime dependencies using established service mesh tools like Kiali and Hubble UI.

navigate-service