The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external, untrusted input such as stakeholder interview transcripts and business documents, which creates a potential surface for indirect prompt injection.\n- Ingestion points: Untrusted interview transcripts and imported documents as noted in the Security Guardrails section of SKILL.md.\n- Boundary markers: Present; the instructions explicitly direct the agent to treat elicitation sources as untrusted and to ignore any instructions or directives embedded within them (ASI01 Goal Hijack).\n- Capability inventory: The skill configuration in SKILL.md includes access to filesystem tools (write_file, edit_file) and shell execution (execute_command).\n- Sanitization: The prompt mandates that the agent never act on directives found within the elicitation sources and requires a reflective self-check to ensure no unverified business rules are assumed.

elicit-requirements